Network Data Envelopment Analysis

International Series in
Operations Research & Management Science
Chiang Kao
Network Data
Envelopment
Analysis
Foundations and Extensions
Second Edition
International Series in Operations Research &
Management Science
Founding Editor
Frederick S. Hillier, Stanford University, Stanford, CA, USA
Volume 340
Series Editor
Camille C. Price, Department of Computer Science, Stephen F. Austin State Uni-
versity, Nacogdoches, TX, USA
Editorial Board Members

Emanuele Borgonovo, Department of Decision Sciences, Bocconi University,
Milan, Italy
Barry L. Nelson, Department of Industrial Engineering & Management Sciences,
Northwestern University, Evanston, IL, USA
Bruce W. Patty, Veritec Solutions, Mill Valley, CA, USA
Michael Pinedo, Stern School of Business, New York University, New York, NY,
USA
Robert J. Vanderbei, Princeton University, Princeton, NJ, USA
Associate Editor
Joe Zhu, Foisie Business School, Worcester Polytechnic Institute, Worcester, MA,
USA
The book series International Series in Operations Research and Management
Science encompasses the various areas of operations research and management
science. Both theoretical and applied books are included. It describes current
advances anywhere in the world that are at the cutting edge of the field. The series
is aimed especially at researchers, advanced graduate students, and sophisticated
practitioners.
The series features three types of books:
• Advanced expository books that extend and unify our understanding of partic-
ular areas.
• Research monographs that make substantial contributions to knowledge.
• Handbooks that define the new state of the art in particular areas. Each
handbook will be edited by a leading authority in the area who will organize a
team of experts on various aspects of the topic to write individual chapters. A
handbook may emphasize expository surveys or completely new advances (either
research or applications) or a combination of both.
The series emphasizes the following four areas:
Mathematical Programming: Including linear programming, integer program-
ming, nonlinear programming, interior point methods, game theory, network opti-
mization models, combinatorics, equilibrium programming, complementarity
theory, multiobjective optimization, dynamic programming, stochastic program-
ming, complexity theory, etc.
Applied Probability: Including queuing theory, simulation, renewal theory,
Brownian motion and diffusion processes, decision analysis, Markov decision
processes, reliability theory, forecasting, other stochastic processes motivated by
applications, etc.
Production and Operations Management: Including inventory theory, produc-
tion scheduling, capacity planning, facility location, supply chain management,
distribution systems, materials requirements planning, just-in-time systems, flexible
manufacturing systems, design of production lines, logistical planning, strategic
issues, etc.
Applications of Operations Research and Management Science: Including
telecommunications, health care, capital budgeting and finance, economics, market-
ing, public policy, military operations research, humanitarian relief and disaster
mitigation, service operations, transportation systems, etc.
This book series is indexed in Scopus.
Chiang Kao
Network Data Envelopment

Analysis
Foundations and Extensions
Second Edition
Chiang Kao
Department of Industrial &
Information Management
National Cheng Kung University
Tainan, Taiwan
ISSN 0884-8289 ISSN 2214-7934 (electronic)

International Series in Operations Research & Management Science
ISBN 978-3-031-27592-0 ISBN 978-3-031-27593-7 (eBook)
https://doi.org/10.1007/978-3-031-27593-7
© The Editor(s) (if applicable) and The Author(s), under exclusive license to Springer Nature Switzerland
AG 2017, 2023
This work is subject to copyright. All rights are solely and exclusively licensed by the Publisher, whether
the whole or part of the material is concerned, specifically the rights of translation, reprinting, reuse of
illustrations, recitation, broadcasting, reproduction on microfilms or in any other physical way, and
transmission or information storage and retrieval, electronic adaptation, computer software, or by
similar or dissimilar methodology now known or hereafter developed.
The use of general descriptive names, registered names, trademarks, service marks, etc. in this publication
does not imply, even in the absence of a specific statement, that such names are exempt from the relevant
protective laws and regulations and therefore free for general use.
The publisher, the authors, and the editors are safe to assume that the advice and information in this
book are believed to be true and accurate at the date of publication. Neither the publisher nor the authors or
the editors give a warranty, expressed or implied, with respect to the material contained herein or for any
errors or omissions that may have been made. The publisher remains neutral with regard to jurisdictional
claims in published maps and institutional affiliations.
This Springer imprint is published by the registered company Springer Nature Switzerland AG
The registered company address is: Gewerbestrasse 11, 6330 Cham, Switzerland
Preface to the First Edition
How to use fewer resources to generate more outputs and services is a concern of all
organizations, including profit pursuing, government, non-profit, and all other types
of decision-making units (DMUs). This is a problem of efficiency, which has three
phases, efficiency measurement, target setting, and goal achievement. Such issues
have been studied by economists and management scientists for many years. Since
the seminal work of Charnes, Cooper, and Rhodes in 1978, Data Envelopment
Analysis (DEA) has become the preeminent nonparametric method for measuring
the efficiency of DMUs that apply multiple inputs to produce multiple outputs. In
addition to efficiency measurement, the DEA technique is also able to show how
much output a DMU can be expected to increase with the current amount of input, or
how much input can be saved while producing the current level of output by simply
increasing its efficiency. In other words, a target for inefficient DMUs to achieve to
become efficient is also provided. The DEA technique is thus able to answer the
questions that arise in the first two phases of efficiency studies. As a consequence,
thousands of papers and dozens of books related to DEA have been published since
its introduction in 1978.
A system is usually composed of many subsystems operating interdependently.
Conventional DEA only considers the inputs supplied to and the outputs produced
from the system in measuring efficiency, ignoring its internal structure. As a result, it
is possible that the overall system is efficient, even while all component divisions are
not. More significantly, there are cases in which all the component divisions of a
DMU have performances that are worse than those of another DMU, and yet the
former still has the better system performance. With an eye on solving these
problems, many ideas have been extended from the conventional DEA to build
models to measure the efficiency of production systems with different network
structures, which are referred to as network DEA. However, these ideas are scattered
in different publications, which are inconvenient to access, and are difficult for
beginners to read, due to a lack of background knowledge. More seriously, some
ideas have already been demonstrated to be incorrect. It is thus desirable to have a
book that presents the underlying theory, model development, and applications of
v
vi Preface to the First Edition
network DEA in a systematic way, to give the readers an idea of what should be done
when developing a new model. It is also desirable to have a book that discusses the
existing models for measuring the efficiency of systems with specific network
structures and explores the relationships between the system and division efficien-
cies. Separating large operations into detailed smaller ones can help identify the
divisions that cause inefficiencies in the system. Novel applications are attractive to
readers in introducing a method and can also inspire further studies. It will thus be
helpful to have a book that describes these applications. Fifteen years after the first
appearance of the term network DEA, there is only one edited book of papers from
different authors on specific topics related to network DEA. A book that has the
above-mentioned functions is thus still needed, and so the current work was written
to meet this need, with the encouragement of Professor Joe Zhu, Editor of Springer’s
International Series in Operations Research and Management Science.
For systems composed of interrelated divisions, managers need to know how the
performances of the various divisions are evaluated and how they are aggregated to
form the overall performance of the system. This book provides an advanced
exposition on evaluating the performance of systems with network structures. It
explores the network nature of most production and operation systems and explains
why network analyses are necessary. The discussion of network DEA carried out in
this work also clarifies the concept of the conventional whole-unit DEA.
In addition to the conventional connected models in the network DEA, this book
highlights a relational model, which is able to show the relationship between the
efficiency of a system and those of its component divisions, when the systems being
examined have different types of network structures. This relationship shows the
extent to which the efficiency of a division impacts that of the system as a whole. The
division with the largest effect is the one that more effort should be devoted to, so
that the performance of the overall system can be raised in a more effective manner.
This book has several features, as follows. Most models are presented with an
associated figure, showing the network structure of the corresponding problem, and
examples are also supplied, which make this book appropriate for class use and self-
study. An extensive bibliography of current research literature on network DEA is
also included, which should be able to inspire researchers to pursue new areas of
work. This book is intended for graduate students who are taking courses or writing
theses on topics related to performance evaluation, DEA, and multi-criteria decision
analysis. It is also suitable for professors and researchers whose research interests are
related to the above-mentioned topics. Experienced practitioners who want to
measure the performance of production, operation, or any other type of DMUs
will also find this book helpful for their work.
In writing this book, many of my former doctoral students helped in providing
suggestions, comments, and proofreading, including Shiuh-Nan Hwang of Ming
Chuan University (Dean of College of Management), Hong-Tau Lee of National
Chin-Yi University of Technology (Vice President), Shiang-Tai Liu of Vanung
University (former Dean of College of Management), Jehn-Yih Wong of Ming
Chuan University (Dean of College of Tourism), Hui-Chin Tang of National
Kaohsiung University of Applied Sciences (Director of Division of Continuing
Preface to the First Edition vii
and Extension Education), Shih-Pin Chen of National Chung Cheng University

(former Head of Department of Business Administration), Ya-Chi Lin of Southern
Taiwan University of Science and Technology (former Head of Department of
Finance), Wen-Kai Hsu of National Kaohsiung Marine University (Dean of College
of Management), Chin-Lu Chyu of Southern Taiwan University of Science and
Technology, Hsi-Tai Hung of Cheng Shiu University, and Pei-Huang Lin of South-
ern Taiwan University of Science and Technology. My assistant Miss Shu-Ting
Hwang drew all the figures and prepared the manuscript according to the format of
the Springer. I am indebted to all of them.
Tainan, Taiwan Chiang Kao

January 2016
Preface to the Second Edition
After the publication of the first edition of this book in 2017, the field of network
DEA has matured considerably. Many new ideas for measuring the efficiency of
network production systems have been proposed. The relationship between the
efficiency of the system as a whole and of its component divisions is being explored,
with the aim of helping the decision maker identify the divisions that cause the
system to perform unsatisfactorily. There are also a number of properties of network
DEA models which despite being widely accepted have been proven to be true only
under certain circumstances. These changes in the field necessitate a revision of the
original book.
In this new edition, the first eight chapters on the measurement of the efficiency of
whole-unit systems have been left mostly unchanged. Elaborations and modifica-
tions to the chapters discussing network systems start from Chap. 9 to which the idea
of using a cooperative model to measure efficiency has been added. The content of
Chap. 14 in the first edition, titled “Hierarchical Systems” has been merged into
Chap. 13, Parallel Systems, because the hierarchical system is a special type of
parallel systems, and it is better to be discussed with parallel systems. The assembly
and disassembly systems which were discussed in the original Chap. 15 are actually
mixed systems. Since they do not have structures that can be expressed in a
systematic way similar to the way in which the series and parallel systems are
expressed, this chapter has been deleted. Chaps. 16, Mixed Systems, and 17,
Dynamic Systems, in the first edition become Chaps. 14 and 15, respectively, in
the second edition.
In addition to these changes, two new chapters have been added: Chap. 16:
“Linkage Efficiency” and Chap. 17: “External and Internal Evaluations.” Conven-
tionally, the intermediate products produced in the system are implicitly assumed to
be fully utilized inside the system. However, recent research suggests that the
intermediate products are not necessarily entirely consumed inside the system,
which leads to the idea of linkage efficiency. This idea and related issues in the
measurement of efficiency are discussed in Chap. 16. Another issue in measuring the
ix
x Preface to the Second Edition
efficiency of a network system is whether the intermediate products should be

considered in defining the efficiency of the system as a whole. This argument is
made clear if we distinguish the viewpoints of the external party and the internal
managers. Chapter 17 is devoted to the discussion of these ideas.
Since the original publication of this book, I have received valuable feedback
from many readers bringing to my attention such potential problems as unclear
descriptions, mistakes, and typos. I am really grateful to them. Their suggestions
have been taken into account while preparing this new edition. As was the case with
the first edition, I am indebted to my assistant, Ms. Shu-Ting Hwang, who prepared
the manuscript according to the formatting requirements of Springer Publishing.
Tainan, Taiwan Chiang Kao

January 2023
Contents
1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.1 History of Network DEA . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.2 Basic Ideas of Efficiency Measurement . . . . . . . . . . . . . . . . . . . 3
1.3 Multi-Input Case . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
1.4 Multi-Output Case . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
1.5 Whole-Unit Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
1.6 Network Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
1.7 Supplementary Literature . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
2 Output–Input Ratio Efficiency Measures . . . . . . . . . . . . . . . . . . . . . 19
2.1 CCR Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
2.1.1 Input Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
2.1.2 Output Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
2.2 BCC Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
2.2.1 Input Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
2.2.2 Output Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
2.3 Restrictions on Multipliers . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
2.4 Ranking . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40
3 Distance Function Efficiency Measures . . . . . . . . . . . . . . . . . . . . . . 43
3.1 Production Possibility Set . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44
3.2 Input Distance Function . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
3.3 Output Distance Function . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54
3.4 Directional Distance Function . . . . . . . . . . . . . . . . . . . . . . . . . 59
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62
xi
xii Contents
4 Slacks-Based Efficiency Measures . . . . . . . . . . . . . . . . . . . . . . . . . . 65

4.1 Additive Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66
4.2 Russell Measures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
4.2.1 Input model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
4.2.2 Output Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72
4.2.3 Input–Output Average Model . . . . . . . . . . . . . . . . . . . 74
4.3 Russell Ratio Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78
4.4 A Classification of Efficiency Measures . . . . . . . . . . . . . . . . . . 82
4.5 Complementary Literature . . . . . . . . . . . . . . . . . . . . . . . . . . . . 85
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 86
5 Efficiency Measurement in Special Production Stages . . . . . . . . . . . 89
5.1 Multiplicative Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 90
5.1.1 Variable Returns to Scale . . . . . . . . . . . . . . . . . . . . . . 91
5.1.2 Constant Returns to Scale . . . . . . . . . . . . . . . . . . . . . . 95
5.2 Free Disposal Hull . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 98
5.2.1 General Case . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 99
5.3 Congestion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 104
5.3.1 Weak Disposability Model . . . . . . . . . . . . . . . . . . . . . 105
5.3.2 Slack-Measure Model . . . . . . . . . . . . . . . . . . . . . . . . . 107
5.3.3 Input-Fixing Model . . . . . . . . . . . . . . . . . . . . . . . . . . 108
5.3.4 Comparison . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 109
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 112
6 Special Types of Input and Output Factors . . . . . . . . . . . . . . . . . . . 115
6.1 Non-discretionary Factors . . . . . . . . . . . . . . . . . . . . . . . . . . . . 116
6.1.1 Input Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 116
6.1.2 Output Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 118
6.1.3 Dual Model Interpretation . . . . . . . . . . . . . . . . . . . . . . 120
6.2 Undesirable Factors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 123
6.2.1 Input–Output Exchange Approach . . . . . . . . . . . . . . . . 123
6.2.2 Data Transformation . . . . . . . . . . . . . . . . . . . . . . . . . . 124
6.2.2.1 Inverse Transformation . . . . . . . . . . . . . . . . 124
6.2.2.2 Negative Transformation . . . . . . . . . . . . . . . 125
6.2.2.3 Shifted Negative Transformation . . . . . . . . . 125
6.2.3 Weak Disposability Approach . . . . . . . . . . . . . . . . . . . 126
6.2.3.1 Hyperbolic Model . . . . . . . . . . . . . . . . . . . . 128
6.2.3.2 Directional Distance Model . . . . . . . . . . . . . 129
6.2.3.3 Variable-Reduction Model . . . . . . . . . . . . . . 130
6.2.4 Slacks-Based Approach . . . . . . . . . . . . . . . . . . . . . . . 131
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 133
Contents xiii
7 Special Types of Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 137

7.1 Negative Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 138
7.2 Ordinal Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 141
7.3 Qualitative Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 142
7.4 Stochastic Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 146
7.5 Interval Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 151
7.6 Fuzzy Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 153
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 159
8 Changes of Efficiency over Time . . . . . . . . . . . . . . . . . . . . . . . . . . . 163
8.1 Theoretic Foundation of MPI . . . . . . . . . . . . . . . . . . . . . . . . . . 164
8.1.1 Input Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 164
8.1.2 Output Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 165
8.1.3 Productivity Index . . . . . . . . . . . . . . . . . . . . . . . . . . . 165
8.2 DEA Measurement of MPI . . . . . . . . . . . . . . . . . . . . . . . . . . . 167
8.3 Global Malmquist Productivity Index . . . . . . . . . . . . . . . . . . . . 173
8.4 Luenberger Productivity Index . . . . . . . . . . . . . . . . . . . . . . . . . 176
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 180
9 Basic Ideas in Efficiency Measurement for Network Systems . . . . . 183
9.1 The Black-Box Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 185
9.2 Independent Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 187
9.2.1 Multiplier Form . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 187
9.2.2 Envelopment Form . . . . . . . . . . . . . . . . . . . . . . . . . . . 188
9.2.3 Slacks-Based Form . . . . . . . . . . . . . . . . . . . . . . . . . . . 189
9.3 Connected Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 190
9.3.2 Multiplier Form . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 193
9.4 Relational Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 195
9.4.1 Multiplier Form . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 196
9.5 Cooperative Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 199
9.5.2 Multiplier Form . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 200
9.6 An Example . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 202
9.6.1 Independent Model . . . . . . . . . . . . . . . . . . . . . . . . . . . 203
9.6.2 Connected Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . 206
9.6.3 Relational Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . 210
9.6.4 Cooperative Model . . . . . . . . . . . . . . . . . . . . . . . . . . . 213
xiv Contents

References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 216
10 Basic Two-Stage Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 219
10.1 Independent Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 220
10.2 Ratio-Form Efficiency Measures . . . . . . . . . . . . . . . . . . . . . . . 223
10.2.1 Efficiency Decomposition . . . . . . . . . . . . . . . . . . . . . . 223
10.2.1.1 Constant Returns to Scale . . . . . . . . . . . . . . 223
10.2.1.2 Variable Returns to Scale . . . . . . . . . . . . . . 226
10.2.1.3 Game-Theoretic Approach . . . . . . . . . . . . . . 230
10.2.2 Efficiency Aggregation . . . . . . . . . . . . . . . . . . . . . . . . 231
10.3 Distance Function Efficiency Measures . . . . . . . . . . . . . . . . . . . 236
10.3.1 System Parameter . . . . . . . . . . . . . . . . . . . . . . . . . . . . 236
10.3.2 Division Parameters . . . . . . . . . . . . . . . . . . . . . . . . . . 240
10.4 Slacks-Based Efficiency Measures . . . . . . . . . . . . . . . . . . . . . . 243
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 246
11 General Two-Stage Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 251
11.1 Feedback System . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 252
11.2 Independent Efficiency Measures . . . . . . . . . . . . . . . . . . . . . . . 258
11.3.1 Game Approach . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 261
11.4.3 Directional Distance Parameter . . . . . . . . . . . . . . . . . . 276
11.6 Shared Input . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 281
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 288
12 Multi-Stage Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 291
12.1 Basic Series Structure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 292
12.2 Independent Efficiency Measures . . . . . . . . . . . . . . . . . . . . . . . 298
Contents xv
12.6 Reversal Links . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 314

12.6.1 Ratio-Form Efficiency Measures . . . . . . . . . . . . . . . . . 315
12.6.2 Slacks-Based Efficiency Measures . . . . . . . . . . . . . . . . 316
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 322
13 Parallel Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 325
13.1 Multi-component Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . 326
13.2 Multi-function Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 331
13.3 Shared Input . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 334
13.3.1 Ratio-form Efficiency Measures . . . . . . . . . . . . . . . . . 335
13.3.2 Distance Function Efficiency Measures . . . . . . . . . . . . 338
13.3.2.1 System Parameter . . . . . . . . . . . . . . . . . . . . 338
13.3.2.2 Division Parameters . . . . . . . . . . . . . . . . . . 340
13.3.2.3 Directional Distance Parameter . . . . . . . . . . 343
13.3.3 Slacks-based Efficiency Measures . . . . . . . . . . . . . . . . 344
13.4 Hierarchical Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 347
13.4.1 Multi-component Systems . . . . . . . . . . . . . . . . . . . . . . 347
13.4.2 Multi-function Systems . . . . . . . . . . . . . . . . . . . . . . . . 352
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 359
14 Mixed Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 363
14.1 Three-Division Structures . . . . . . . . . . . . . . . . . . . . . . . . . . . . 364
14.1.1 The R&D Value Chain Example . . . . . . . . . . . . . . . . . 364
14.1.2 The Railway Operation Example . . . . . . . . . . . . . . . . . 367
14.1.3 The Transportation Network Example . . . . . . . . . . . . . 370
14.1.4 The Bank Profit Centers Example . . . . . . . . . . . . . . . . 372
14.2 Four-Division Structures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 375
14.2.1 The International Tourist Hotel Example . . . . . . . . . . . 375
14.2.2 The Environmental Protection Example . . . . . . . . . . . . 377
14.2.3 The Corporate and Consumer Banking Example . . . . . . 380
14.2.4 The Matrix Structure Example . . . . . . . . . . . . . . . . . . 383
14.3 Five-Division Structures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 385
14.3.1 The Major League Baseball Example . . . . . . . . . . . . . . 385
14.3.2 The NBA Basketball Example . . . . . . . . . . . . . . . . . . . 389
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 392
15 Dynamic Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 395
15.1.1 The Whole-unit Case . . . . . . . . . . . . . . . . . . . . . . . . . 397
15.1.2 The Network Case . . . . . . . . . . . . . . . . . . . . . . . . . . . 398
15.2.1 The Production Delays Example . . . . . . . . . . . . . . . . . 401
15.2.2 The Period Distance Parameters Case . . . . . . . . . . . . . 402
xvi Contents
15.2.3 Directional Distance Function: Whole Unit . . . . . . . . . 403

15.2.4 Directional Distance Function: Network . . . . . . . . . . . . 405
15.3.1 The Basic Dynamic Structure Case . . . . . . . . . . . . . . . 407
15.3.2 The Aggregate Slack Case . . . . . . . . . . . . . . . . . . . . . 408
15.3.3 The Network Case . . . . . . . . . . . . . . . . . . . . . . . . . . . 410
15.4 Value-Based Efficiency Measures . . . . . . . . . . . . . . . . . . . . . . 412
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 416
16 Linkage Efficiency . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 419
16.1 Basic Series Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 420
16.1.1 Two-division Systems . . . . . . . . . . . . . . . . . . . . . . . . 420
16.1.2 Multi-division Systems . . . . . . . . . . . . . . . . . . . . . . . . 424
16.2 Parallel Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 429
16.3 General Network Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . 432
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 445
17 External and Internal Evaluations . . . . . . . . . . . . . . . . . . . . . . . . . 447
17.1 External Evaluation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 448
17.1.1 Radial Efficiency . . . . . . . . . . . . . . . . . . . . . . . . . . . . 449
17.1.2 SBM Efficiency . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 453
17.2 Internal Evaluation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 454
17.2.1 Radial Efficiency . . . . . . . . . . . . . . . . . . . . . . . . . . . . 455
17.2.2 SBM Efficiency . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 457
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 462
18 Epilogue . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 465
18.1 Generality of Some Representative Models . . . . . . . . . . . . . . . . 465
18.2 Which Model to Use . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 469
18.3 Road Ahead . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 471
Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 473
About the Author
Chiang Kao is a University Chair Professor of Industrial and Information Manage-

ment at National Cheng Kung University in Taiwan. He received his MS in
Operations Research and PhD in Forest Management from Oregon State University
and did post-doctoral research in Economics also at Oregon State University. After
joining Cheng Kung University, he has taken sabbatical leaves at Purdue University,
Aachen University of Technology (Germany), INSEAD and University of Paul
Sabatier (France), and University of Bologna (Italy). He has also been hired by the
Computer Science Department of Southwest Texas State University.
Professor Kao has also held administrative posts at Cheng Kung University,
including Chairman of Department of Industrial and Information Management,
Dean of College of Management, and President of the University. He has published
more than eighty papers on DEA in international journals, of which more than thirty
are related to network DEA. Currently, he serves as a member of the Editorial Board
for European Journal of Operational Research and an Associate Editor for Omega.
xvii
Chapter 1
Introduction
Performance evaluation is an important task in management, to better understand the

past accomplishments of a production unit and plan for its future development. The
objective is to understand either how far this unit can be expected to increase its
output with the current amount of input or how much input can be saved while
producing the current level of output by simply increasing its efficiency.
For profit-pursuing organizations the evaluation is relatively easy, as monetary
terms such as revenues, returns on investment, and net profit can be used to
aggregate inputs and outputs together to form an index to show whether an organi-
zation is making a profit or not. However, if some inputs or outputs do not have a
market value, or if profit is not the only goal of the organization, then calculating
these indices becomes difficult. This task is even more difficult for not-for-profit
organizations, where most of the services provided do not have a market value,
which makes an aggregation of the inputs used and the outputs (including services)
produced difficult. In this case one usually looks at ratios of one output to one input
of specific interest. If a measure of the overall performance is desired, then all ratios
must be aggregated to form one index. For example, a forest uses land, labor,
expenses, and forest stocks to produce timber, conserve soil, absorb carbon dioxide
and release oxygen, and provide a habitat for wildlife and scenery for recreation. To
measure the performance of a forest, all combinations of the output–input ratios, e.g.,
timber produced per employee and carbon dioxide absorbed per unit area, are
calculated. Since the ratios are non-commensurable, they need to be standardized,
and suitable weights must be found to aggregate them to form one index that can be
compared with that calculated from the standards. Different ways of standardization
and different weights for aggregation obviously result in different indexes. This
approach is thus difficult to use in practice. Many economists and management
scientists have thus been working to develop a suitable way to objectively measure
the performance of a production unit that uses multiple inputs to produce multiple
outputs.
© The Author(s), under exclusive license to Springer Nature Switzerland AG 2023 1

C. Kao, Network Data Envelopment Analysis, International Series in Operations
Research & Management Science 340,
https://doi.org/10.1007/978-3-031-27593-7_1
2 1 Introduction
1.1 History of Network DEA
In physical sciences theoretical standards showing the relationship between several

factors often exist, which makes the performance measurement of a production unit
possible. However, this is not the case for social sciences, where empirical standards
compiled from observations must be relied on. The work of Farrell (1957) on the
productive efficiencies of agriculture in forty-eight states in the USA was probably
the first empirical study on efficiency measurement. This work constructed isoquants
as the standards and then used these to measure the productive efficiency. In this
pioneering work the number of outputs was limited to one, and the returns to scale
were assumed to be constant. These were then relaxed to be variable by Farrell and
Fieldhouse (1962). One characteristic of this method is that it does not need to
specify the mathematical form of the production function, and it is thus referred to as
a nonparametric approach. In contrast to this, Aigner and Chu (1968) adopted a
Cobb-Douglas production function and applied the mathematical programming
technique to estimate the parameters in this for the primary metals industry in the
USA, with the number of outputs also limited to one. Since the mathematical form of
the production function is known, this is referred to as a parametric approach.
Charnes et al. (1978) proposed a fractional programming technique for aggregat-
ing the outputs into a virtual output and the inputs into a virtual input and used the
ratio of these to represent the relative efficiency of a production unit (generalized as a
decision making unit, DMU, in their study). In its dual formulation, the observations
are enveloped by the production function, and the technique was thus named data
envelopment analysis, abbreviated as DEA. It was in the doctoral dissertation of the
third author of this study (Rhodes, 1978) that this name was first introduced. Since
the function form need not be specified, it is a nonparametric approach. The returns
to scale were assumed to be constant in Charnes et al. (1978). In a later study by
Banker et al. (1984), this assumption was relaxed to allow for variable returns to
scale. In the same year Byrnes et al. (1984) derived the same measure of efficiency
based on the concept of distance functions (Shephard, 1970). Førsund et al. (1980)
classified the methods for measuring efficiencies into six categories, although they
basically belong to two approaches, parametric and nonparametric, as noted above.
After the work of Charnes et al. (1978), numerous studies on methodologies and
applications of DEA have been reported. Indeed, this number has been increasing
year by year, as shown in several review studies (Seiford, 1996; Gattoufi et al., 2004;
Emrouznejad et al., 2008; Cook & Seiford, 2009; Liu et al., 2013a, b; Lampe &
Hilgers, 2015; Panwar et al., 2022). Such studies are concentrated on measuring the
efficiency of DMUs as a whole unit, in that only the inputs supplied to and the
outputs produced by the DMU are considered. In reality, organizations are often
comprised of a number of divisions, each of which performs different functions,
interacting in a multitude ways to give rise to organizational performance. When
interactions among divisions are not taken into account, the results will be distorted
and misleading. For example, the operations of banking and similar industries have
two processes, capital collection and investment. If one studies the impact of
1.2 Basic Ideas of Efficiency Measurement 3
information technology (IT) on the performance of a bank, it may be found that it has
no direct effect on profitability. This is because IT is useful for capital collection,
while whether a bank would make a profit or not is dependent on the correct
investment decisions being made. Therefore, to study the performance of a DMU
it is necessary to study its component divisions, so that the cause of any inefficien-
cies can be identified, and the measured efficiency will be meaningful.
In the early stages of the development of DEA, Charnes et al. (1986) noticed this
phenomenon in studying army recruitment, which had two processes: the first
created awareness through advertisements, and the second created contracts. Sepa-
rating large operations into detailed smaller ones helps identify the real impact of
input factors. Although the idea of taking the component divisions of a production
system into consideration in measuring efficiencies had been discussed in the
literature in the 1980s and 1990s, the first appearance of the term network DEA
was in 2000 (Färe & Grosskopf, 2000). The whole-unit, or black-box, performance
measurement is relatively simple to conduct, because only the inputs supplied to and
the outputs produced by the DMU need to be considered, which makes a systematic
expression of the model possible. The network system performance measurement, in
contrast, is difficult to express using a general model, because different structures of
the network production system are involved. The simplest structure of network
systems is the two-stage tandem system, where the inputs supplied from outside
are all used in the first stage to produce a set of intermediate products to be used in
the second stage to produce the final outputs of the system. Due to its simple
structure and wide application, this system has been extensively discussed in the
literature [see the review of Cook et al. (2010)]. However, as a system becomes more
complicated, the systematic expression of the model is only possible for certain
specific types of system, and for other unstructured systems, this remains difficult.
Along with the other problems and issues that already exist with regard to whole-unit
analyses, this can make network analysis even more complicated. The review
presented in Kao (2014) with regard to network DEA presents details of many
related models and applications and gives a good overview of this topic.
The purpose of this book is to introduce the ideas underlying network DEA, the
models that have been developed to measure the efficiency of network production
systems, and the applications that have been made in the real world. To aid in this the
basic concepts of DEA, which network DEA is based on will also be discussed.
1.2 Basic Ideas of Efficiency Measurement
In economic theory a production function is a mathematical equation showing the

maximum amount of output that can be produced from any specified set of inputs,
given the existing technology (Ferguson & Gould, 1986). In this context, the
production function sets the highest possible limit on the output which a production
unit can hope to obtain with a certain combination of inputs. It is the highest level
because the inputs of a production unit may not have been perfectly utilized, such
4 1 Introduction
Fig. 1.1 Production Y

functions and efficiency
measures 15
Q
YV V D
10 YH F H R
C
YA U P
A
YG J
G
5
B XU XA
XJ
O 5 10 15 X
that its production is less than what is expected. The difference between the
maximum amount of output and the actual amount produced by a production unit
is due to the inefficiency of this unit.
The production function is usually not known, and it has to be estimated from the
data collected from real life. Since the sampling process used to achieve this may not
be exhaustive, and it is also possible that none of the units is operating in an efficient
manner, the empirically estimated production function always lies below the true
one. The efficiency of a unit measured from this underestimated production function
is thus always overstated. For this reason, the efficiency measured from the empirical
production function is called relative efficiency, as opposed to absolute efficiency,
highlighting the fact that it is a measure that exists in relation to the other sampled
units.
Consider a simple production process, where one-input X is used to produce one
output Y. Ten units with the (X, Y ) observations of (2, 1), (4, 1), (4, 2), (5, 3), (6, 6),
(8, 6), (8, 9), (10, 8), (12, 10), and (14, 7) are collected and are depicted in Fig. 1.1.
Let the production function be represented by Y = ea - b/X, or lnY = a - b/X, where
a and b are parameters to be estimated. The conventional regression analysis for
minimizing the total squared logarithmic distance: min 10 i = 1 ða - b=X i - ln Y i Þ
2
yields a = 2.5093 and b = 5.9807, with the regression curve shown by OP in

Fig. 1.1. The reason for using logarithmic distance here is to make the regression
model linear. If linearity is not of concern, then one can use the nonlinear form of
a - b=X i 2
minimizing 10 i=1 e - Y i to find the normal distance. Curve OP shows the
average amount of output that can be produced from a given amount of input.
Theoretically, one-half of the units lie above this curve and the other half lie
below it. Let Y i denote the value on this average production curve with input Xi.
The ratio of Yi to Y i serves as a measure of performance for a point i, where a value
greater than unity indicates that the corresponding unit performs better than average,
1.2 Basic Ideas of Efficiency Measurement 5
and a value less than unity indicates the opposite. Larger measures basically imply
better performance. This measure shows the relative performance of a unit, but it
does not show how far a unit is from the maximum amount of output that it can
produce. Even if a unit has the largest measure among all units, it does not mean that
it has reached the maximum amount of production. It is possible that one unit has a
measure larger than another unit, yet lies relatively farther below the maximum
output.
To find the production function of maximum output, one must find a curve lying
above all observations to reflect the fact that it is the maximum output for any given
amount of input. Regression analysis can still be used to find such a curve, where the
total squared distance between the actual and theoretical output (the output on the
curve) for all sampled units is minimized, while additionally requiring the observed
output to be less than or equal to the corresponding output on the curve for each unit.
This idea can be represented by a quadratic programming model with the
following form:
10
2
min: ða - b
Xi - ln Y i Þ
i=1
b ð1:1Þ
s:t: a - ≥ ln Y i , i = 1, . . . , 10
Xi
a, b ≥ 0
which has a solution of a = 2.9296 and b = 5.8593. Curve OQ in Fig. 1.1 shows this
production function. Since the production function always lies above all observa-
tions, it is also called the production frontier. The efficiency of a unit is the ratio of its
actual output to that on the production frontier, such that it is always less than or
equal to unity. For example, the efficiency of unit A is EA = YA/YV. The performance
index measured from the average output curve OP, as obtained from the regression
analysis, is YA/YG, which is greater than unity, indicating that this unit performs
better than average. However, this unit is still inefficient, because its output is less
than the maximum amount that can be produced.
The ratio of YA to YV is the efficiency of unit A measured from the output side, in
that the maximum output under a specified amount of input is desired. The efficiency
can also be measured from the input side, with the minimum input that is able to
produce the specified amount of output is of interest here. In Fig. 1.1, XU is the
smallest input level that is able to produce the same output level of unit A. In other
words, the input level of unit A can theoretically be reduced from XA to XU, with the
same output level of YA being produced. XU/XA is thus the efficiency of unit A from
the input side. The efficiency measured from the output side need not be the same as
that measured from the input side.
In this example the mathematical form of the production function for both the
average amount and the maximum amount cases has been specified, with the
parameters a and b left to be determined from the sample. Obviously, mistakenly
specified function forms lead to erroneous efficiency measures. From the
6 1 Introduction
observations in Fig. 1.1, it is clear that the piecewise line segments connecting the
outermost observations superimposing upon all observations, BCDR, are the most
favorable production frontier, in terms of the efficiency scores measured for all units.
In this example, the output-oriented efficiency for unit A is YA/YH, which is greater
than that measured from the production frontier OQ, YA/YV, and the input-oriented
efficiency is XJ/XA, which is also greater than that measured from the production
frontier OQ, XU/XA. Since the piecewise linear production frontier BCDR envelops
all observations, the associated technique for determining the frontier and subse-
quently measuring efficiencies is named data envelopment analysis (abbreviated as
DEA). This technique does not need to specify the function form of the production
function, and thus it is a nonparametric approach.
The parametric production frontier OQ is constructed by minimizing the distance
between the actual output and the maximum output on the curve. This curve can be
regarded as an output-oriented one. Measuring the input efficiency based on an
output-oriented production frontier will in general understate the efficiency. To give
a fairer input efficiency measure for the unit, the production frontier should be
constructed from the input side; that is, to minimize the distance between the actual
input and the corresponding input on the curve with the same output level. Specif-
ically, the production function with output Y as the independent variable, lnY = a -
b/X, should be expressed as 1/X = a/b - (lnY )/b = a′ - b′ ln Y, with input X as the
dependent variable. The right-hand side of the equation is the minimum input on the
frontier. The associated quadratic programming model is used to minimize the total
squared distance between the minimum input and the actual input:
10 0 0 2 ′ ′
i = 1 ða - b ln Y i - 1=X i Þ , under the constraints of a - b ln Yi ≤ 1/Xi, i = 1,
. . ., 10. It is noted that the distance in this case is measured in the reciprocal form of
1/X, to make the model linear. The results from the input side are a = 2.7243 and
b = 10.8971, which are different from those obtained from the output side; however,
the former approach gives a better measure of input efficiencies. While the paramet-
ric approach obtains different production frontiers from the input and output sides,
the nonparametric DEA approach yields the same production frontier, because the
outermost observations are the same, no matter whether they are viewed from the
output or the input side.
The input-oriented model for measuring efficiency is based on the concept of
producing the same amount of output by using the least amount of input. In contrast,
the output-oriented model is based on the concept of consuming the same amount of
input to produce the largest amount of output. In fact, the input and output need not
be considered independently in measuring efficiencies. Referring to Fig. 1.1, any
point on the segment UV of curve OQ (or the piecewise line segments JCH) that
dominates unit A can be used as the benchmark for measuring the efficiency of unit
A. Suppose point F is selected as the benchmark. Then the ratio of the output of unit
A to that of point F is the output efficiency and the ratio of the input of point F to that
of unit A is the input efficiency of unit A. An approach for aggregating the input and
output efficiencies to form the final efficiency is then applied. This approach can be
referred to as the input–output combined approach, and several methods have been
proposed to achieve this.
1.3 Multi-Input Case 7
1.3 Multi-Input Case
The case of using a single input and single output for measuring efficiencies is
simple, no matter whether the input or output side is considered. However, the
problem becomes a little more complicated when the number of inputs is more
than one.
Suppose m inputs, X1, X2, . . ., Xm, are used to produce one output Y, and the
production frontier is described by Y = G(X1, X2, . . ., Xm). In this case the output
efficiency is the ratio of the actual output produced to the maximum output on the
production frontier, which is quite simple to measure. The input efficiency, however,
is not so simple to measure, because there is more than one input involved. Consider
the simplest case of two inputs, as cases of more inputs can be generalized in a
straightforward manner.
Figure 1.2 shows a production frontier, where unit A uses X1A of input 1 and X2A
of input 2 to produce YA of output Y. Its output efficiency is the ratio of YA to YZ,
where YZ = G(X1A, X2A) is the output on the production frontier corresponding to the
input level of (X1A, X2A). To measure its input YA efficiency, we slice the production
frontier horizontally at the level of YA to form an isoquant of YA = G(X1, X2), which
shows all combinations of X1 and X2 that are able to yield the output level YA. We
Fig. 1.2 A production Y

function of two inputs and
one output YZ
YA Z
S
A
S'
O X2
S
(X1A, X2A)
S'
X1
8 1 Introduction
Fig. 1.3 Isoquant and input X2

efficiencies
S R
X2A U A
E H
cA
X2B B
D F
cB Q R'
cE V S'
O X1
X1B X1A
then project this isoquant, denoted as SS′, onto the X1-X2 plane at Y = 0, and redraw
it in Fig. 1.3 to provide a more detailed explanation of this. The points on the
isoquant SS′ are efficient, because they lie on the production frontier. The interior
points of the region to the top right of the isoquant SS′ are inefficient. Theoretically,
any point on the UV segment of the isoquant can be selected as the benchmark for
unit A to measure its efficiency, because all these points use less amounts of X1 and
X2 to produce the same amount of output YA. For example, when unit F on the
isoquant is selected, the X1- and X2-efficiency of unit A are X1F/X1A and X2F/X2A,
respectively. Since these two measures usually are different, a method to aggregate
them to form one index is necessary, which requires a method with persuasive
explanation. However, if we select point B, the intersection of ray OA and the
isoquant SS′, as the benchmark, then we have X1-efficiency, X1B/X1A, being equal
to X2-efficiency, X2B/X2A, which avoids the confusion associated with aggregating
the X1- and X2-efficiency. This measure is called the radial measure.
The isoquant shows all possible combinations of inputs capable of producing a
given level of output. Since different inputs have different market prices, an appro-
priate allocation of the two inputs will produce the same amount of output, while
incurring the least cost. Let c1 and c2 be the unit costs of inputs 1 and 2, respectively.
Then c1X1 + c2X2 = c0 is an isocost curve, which shows different combinations of
inputs 1 and 2 that have a total cost of c0. The isocost curve c1X1 + c2X2 = cE, which
is tangent to the isoquant SS′ at point E, shows that the minimum cost for producing
YA amount of output is cE, and X1E and X2E are the optimal amounts of inputs 1 and
2, respectively, to achieve this minimum cost. Point B has a total cost of
c1X1B + c2X2B = cB. Although this point is efficient from the production point of
view, it costs more than point E does to produce YA level of the output. Unit B thus
has a price efficiency, or allocative efficiency, of cE/cB, which is equal to OD/OB.
1.4 Multi-Output Case 9
Similarly, the total cost associated with unit A is c1X1A + c2X2A = cA. By taking the
price effect into consideration, the efficiency score of unit A becomes cE/cA (which is
equal to OD/OA), where the portion of cE/cB (which is equal to OD/OB) is due to
inappropriate allocation of the inputs and the portion of cB/cA (which is equal to OB/
OA) is due to inefficiency in production. The overall efficiency, OD/OA, is the
product of the allocative efficiency, OD/OB, and the productive efficiency, OB/OA.
The smooth isoquant SS′ of the form YA = G(X1, X2) is obtained from a parametric
approach. When the nonparametric DEA approach is applied, the isoquant obtained
will be some piecewise line segments, such as RPQR′ shown in Fig. 1.3. The
efficiency score for unit A in this case is OH/OA, which is greater than that measured
from the smooth isoquant SS′. The idea of radial measures can be extended to higher
dimensional spaces. In this case, the frontier line segments of the DEA approach
become frontier facets.
1.4 Multi-Output Case
In contrast to using multiple inputs to produce single output, another situation is

applying single input to obtain multiple outputs. Let H(X1, Y1, . . ., Ys) = 0 be the
implicit form of the production frontier. Figure 1.4 shows a production frontier for
the simplest case of s = 2, where one input X is used to produce two outputs, Y1 and
Y2. Units A and Z apply XA and XZ amounts of the input, respectively, to produce the
Fig. 1.4 A production O Y2

function of one input and T
two outputs
(Y1A, Y2A)
T''
XZ
XA T Y1
X T'
10 1 Introduction
Fig. 1.5 Product Y2

transformation curve and
output efficiencies
T
D
U
W P B
H F
Y2A
A V
E
rB
rA
Q
rE
W' T'
O Y1
Y1A
same amounts of Y1 and Y2, where A lies below the production frontier, while Z lies
on the frontier. The input efficiency of unit A is thus XZ/XA. To investigate the output
efficiency of unit A, we slice the production frontier horizontally at the input amount
of XA and project it onto the Y1-Y2 plane at X = 0 and redraw it in Fig. 1.5.
The curve TT′ in Fig. 1.5, projected from the production frontier at the input
amount of XA, can be expressed as XA = H(Y1, Y2). It is a product transformation
curve, which is the locus of output combinations of the amounts of Y1 and Y2 that can
be produced from the input amount XA. All points on the segment UV of this curve
dominate unit A, because they use the same amount of input XA to produce more
amounts of both Y1 and Y2. Any point F on this segment can theoretically be selected
as a benchmark to evaluate the output efficiency of unit A, with Y1- and Y2-efficiency
of Y1A/Y1F and Y2A/Y2F, respectively. However, the problem is how to aggregate
these two efficiencies to form one index. Similar to the multi-input case, one can use
the radial measure by selecting unit B, which is the intersection point of the ray
emanating from the origin and passing through unit A and the product transformation
curve TT′, as the benchmark to result in the same measure of Y1- and Y2-efficiency.
Other ideas for selecting the benchmark from the dominating set UV have been
proposed in the literature, and they will be discussed in succeeding chapters.
Different outputs have different market prices. By utilizing the difference in price,
one can find a point on the product transformation curve that generates the largest
revenue. Let p1 and p2 be the unit prices of outputs 1 and 2, respectively. Unit A has a
total revenue of p1Y1A + p2Y2A = rA, and the linear equation p1Y1 + p2Y2 = rA is an
isorevenue line. This line is the revenue counterpart of an isocost line and is defined
as the locus of output combinations that will earn a specified revenue. Clearly, for
those points on the product transformation curve, point E has the largest revenue,
1.5 Whole-Unit Analysis 11
indicated as rE. Point B generates a lower revenue of rB, although it is also on the
product transformation curve. Its efficiency score can be defined as rB/rE, and the
inefficiency is due to the inappropriate proportions of outputs Y1 and Y2 that are
produced. Similarly, A has an overall efficiency of rA/rE = (OA/OD), where the
portion of rA/rB = (OA/OB) is due to inefficiency in production, usually referred to as
productive efficiency, and the portion of rB/rE = (OB/OD) is due to the inappropriate
proportions of the outputs produced, usually referred to as allocative, or price,
efficiency. The overall efficiency is a product of the productive and allocative
efficiencies.
Similar to the multi-input case, the product transformation curve TT′ of the form
XA = H(Y1, Y2) is obtained from a parametric approach. When the nonparametric
approach, DEA, is applied, the product transformation curve will have a piecewise
line segments form, such as WPQW′. The productive efficiency of unit A in this case
becomes OA/OH, which is greater than that measured from the parametric approach,
OA/OB. The case of two outputs discussed in this section can be generalized to cases
of multiple outputs.
1.5 Whole-Unit Analysis
The most general case of a production process is that multiple inputs are used to
produce multiple outputs. This is very common in the real world. For example, as
stated at the beginning of this chapter, a forest uses land, labor, expenses, and forest
stocks to produce timber, and to provide the services of soil conservation, oxygen
production, carbon dioxide absorption, recreational facilities, and wildlife habitat. A
general production system can be represented by a black box, as depicted in Fig. 1.6,
where m inputs Xi, i = 1, . . ., m are applied to the system to produce s outputs Yr,
r = 1, . . ., s. The relationship between the inputs and outputs can be described by an
implicit form of the production frontier P(X1, . . ., Xm, Y1, . . ., Ys) = 0, in that all units
satisfy the condition of P() ≤ 0. A unit k lying on the production frontier with
P(X1, . . ., Xm, Y1, . . ., Ys) = 0 is efficient, whereas one lying below it, with P(X1, . . .,
Xm, Y1, . . ., Ys) < 0, is inefficient.
There are different ways of constructing the production frontier P(), and different
points on the production frontier can be used as the benchmark to produce different
measures of efficiency. The graphical illustrations in the preceding sections show
that the nonparametric DEA approach yields the most favorable efficiency measures
for all the units being evaluated. For this reason, DEA has been considered an
effective approach for efficiency measurement. Since the production frontier is
constructed from sampled units, the measured efficiency for each unit is a relative
Fig. 1.6 Whole-unit

representation of the Xi Production Yr
production system
i=1,…, m System r=1,…, s
12 1 Introduction
measure, in that the results are relative to other units used to construct the frontier.
When more units are included, the production frontier will be raised, and the
measured efficiency will consequently decrease.
A system is usually composed of many divisions, each performing specific
functions. When one is interested in the performance of the system as a whole
unit, in that only the inputs supplied to and the outputs produced from the system are
considered, it is referred to as a whole-unit, or black-box analysis, because how the
outputs are converted from the inputs via the intermediate products is not a concern.
If the efficiency of a division is of interest, then the division can be treated as a black
box, and only the inputs supplied to and the outputs produced from it are considered.
In other words, the division is considered as a DMU. Suppose the efficiency of a set
of n DMUs is to be evaluated. The jth DMU, j = 1, . . ., n, applies inputs Xij, i = 1,
. . ., m, to produce outputs Yrj, r = 1, . . ., s. Conventionally, efficiencies are measured
from two points of view, input and output, where the former concerns the amount by
which the input can be reduced while still producing the same amount of output, and
the latter concerns the amount by which the output can be increased while still
consuming the same amount of inputs. It was then found that neither the inputs nor
the outputs need to be fixed in measuring efficiencies. Any dominating unit with a
greater amount of outputs and less inputs can be used as the benchmark to measure
efficiency. Models of this type are referred to as combined input–output models. All
of the existing models belong to one of these three types, input, output, and
combined input–output.
The models for measuring efficiencies have different mathematical forms, based
on different concepts of measurement. The first is an output–input ratio measure,
which expresses efficiency as a ratio of the aggregate output to the aggregate input.
The second is a distance function measure, which uses the distance between the
DMU being evaluated and a benchmark on the production frontier to measure
efficiency. The third is slacks-based measure, which uses the excessive inputs
consumed and the shortfalls in outputs to measure efficiency. Chapters 2, 3, and 4
discuss the ideas and models of these measures.
In the classical production theory, production is separated into different stages as
the marginal product of the variable input is increasing, decreasing, and becoming
negative, and different DEA models are developed to evaluate the efficiency of
DMUs at different stages. In addition to production stages, the input and output
factors may have special characteristics, such as whether they are desirable or not,
and the data may have different types, e.g., precise, probabilistic, ordinal, qualitative,
interval, and fuzzy. Chapters 5, 6, and 7 introduce the models for measuring the
efficiency of DMUs at different production stages, those with special factor types,
and those with special data types, respectively. For each DMU, sometimes its
performance over time is of interest, especially due to the effect of an act or policy.
Chapter 8 gives an overall view of the shift of the production frontiers over time. In
this case, there is no connection between two consecutive periods.
1.6 Network Analysis 13
1.6 Network Analysis
The whole-unit analysis gives a general idea of the performance of a unit. However,
since a system is usually composed of several divisions operating interdependently,
ignoring the operations of the component divisions may obtain misleading results. A
number of examples have been presented in the literature to show that an overall
system may be efficient, even while all component divisions are not (Kao & Hwang,
2008). More significantly, there are cases in which all the component divisions of a
unit have performances that are worse than or equal to those of another unit, and yet
the former still has the better system performance (Kao & Hwang, 2010). These
findings indicate that the internal structure of a production system should be con-
sidered, whenever the data is available, to produce correct results when measuring
efficiencies. Figure 1.7 is a conceptual diagram, showing that the component
divisions of a system convert the inputs supplied to the system to the final outputs
via some intermediate products produced and consumed in the system. Many models
have been developed to measure the efficiencies of this type of system, with the
internal structure being considered.
Suppose a system is composed of p divisions. It should be noted that other terms
for this, such as subunits, sub-DMUs, and components, have also been used, and we
ðk Þ ðk Þ
will use the term divisions when there is no ambiguity. Denote X ij and Y rj as the ith
input supplied from outside, i 2 I(k), where I(k) is the index set of the exogenous
inputs used by division k, and the rth final output of the system, r 2 O(k), where O(k)
is the index set of the final outputs produced by division k, k = 1, . . ., p, respectively,
ðk Þ ðk Þ
of the jth DMU. Clearly, the sums of X ij and Y rj for all p divisions are the system
p ðk Þ
input Xij and system output Yrj, respectively, i.e., k = 1 X ij = X ij and
p ðk Þ ða,kÞ
Y
k = 1 rj = Y rj . Further, let Z fj denote the fth intermediate product produced by
division a, j 2 M , where M is the index set of the intermediate products used by
(k) (k)
ðk,bÞ
division k, and Z gj denotes the gth intermediate product to be used by division
b, g 2 N , where N(k) is the index set of the intermediate products produced by
(k)
division k. The same intermediate product f produced by different divisions for

ðk Þ ða,k Þ ðk Þ
division k to use can be aggregated as Z fj , i.e., pa = 1 Z fj = Z fj . Similarly, the
same intermediate product g produced by division k for different divisions to use can
ðk Þ ðk,bÞ
be aggregated as Z gj = pb = 1 Z gj . With these notations, the general network
structure can be depicted, as shown in Fig. 1.8.
Fig. 1.7 Network

representation of the Division 1 Division p
production system
Xi ... Yr
i=1,…, m Division 2 r=1,…, s
Production System
14 1 Introduction
Fig. 1.8 The operation of X i( k ) , i ∈ I ( k )

each division k in a network
production system
Z (f1,k ) Z g(k ,1)

Z (f 2,k ) Z (k
f
)
Z (k )
g Z g(k , 2)
k
f ∈ M (k ) g ∈ N (k )
…
…
Z (f p ,k ) Z g( k , p )
Yr( k ) , r ∈ O ( k )
There are two basic concepts being used in developing network DEA models,
efficiency decomposition and efficiency aggregation. The basic idea of efficiency
decomposition is to measure the efficiency of a system based on the inputs supplied
to it and the outputs produced, while taking the operations of the component
divisions into consideration. The efficiency of each division is measured based on
the inputs consumed, both exogenous and intermediate, and the outputs produced,
both final and intermediate, by the division. The efficiency of the system can usually
be decomposed into those of the component divisions, expressed by a function. For
efficiency aggregation, the efficiency of the system is defined as a function of those
of the divisions. In this case, the intermediate products are included in calculating the
efficiency of the system, because they have been included in measuring the effi-
ciency of the division, and this is the main difference between this approach and the
efficiency decomposition approach.
Network production systems have different structures, and the discussion of
efficiency measurement is usually based on the structure of the system. The simplest
structure of the network system, and the one which has been discussed the most, is
the two-division tandem system, usually called the two-stage system, where all the
inputs are supplied to the first division to produce a number of intermediate products
for the second division to produce all the final outputs. An extension of this structure
is the general two-stage system, where some of the outputs of the first division are
sent out of the system to become final outputs, and the second division also
consumes inputs supplied from outside. The general two-stage system can be
generalized to multiple stages. All these systems have a series structure. Another
structure, as opposed to series, is the parallel structure, where a number of divisions
operating independently in the system, each consumes a number of inputs supplied
from outside and produces a number of final outputs. All these structures can be
described by models in a systematic way.
In addition to these structured systems, there is a system called hierarchy system,
where each division has several subordinated divisions associated with from the top
level to the bottom level. The inputs in a division are transmitted to its subordinated
divisions, without operations. Although the hierarchy system has a clear structure of
1.7 Supplementary Literature 15
levels of divisions, the number of divisions and the number of levels in each branch
of the hierarchy structure are not necessarily the same, which makes a systematic
way of modeling the system impossible. However, since this type of system can be
represented by a parallel system composed of the divisions at the bottom of the
hierarchical structure, it is thus discussed in the chapter of Parallel Systems.
Chapter 9 will discuss the basic idea of efficiency measurement for general network
systems. Chapters 10–14 will then discuss the two-stage tandem system, general
two-stage system, general multi-stage system, parallel system, and mixed system
(a general term for non-structured systems), in sequence. For each structure the
efficiency is measured in a static manner. When time is involved, then one has a
dynamic system. Chapter 15 discusses efficiency measurement for network systems
over time.
Chapter 16 introduces the concept of linkage efficiency, which is a measure of
how much percent of an intermediate product produced in the system is being
utilized. Chapter 17 discusses how efficiencies are measured from the viewpoints
of outside peers and managers of the production system. The discussion concentrates
on the decomposition of the efficiency of the system to the efficiencies of the
divisions and the aggregation of the efficiencies of the divisions to form the
efficiency of the system. Finally, Chap. 18, Epilogue, discusses what model to use
when faced with a problem, and miscellaneous problems in network DEA that need
further research so that more real-life problems can be solved, and more managerial
information can be provided.
1.7 Supplementary Literature
DEA has been widely applied to measure the efficiency of organizations in the real
world. In addition to the general review articles mentioned in Sect. 1.1, there are
many other reviews of DEA applications in specific areas. They include financial
institutions (Berger & Humphrey, 1997; Fethi & Pasiouras, 2010; Paradi & Zhu,
2013), energy and environmental studies (Zhou et al., 2008; Song et al., 2012; Xu
et al., 2020; Chachuli et al., 2020), agriculture (Balezentis, 2014), aquaculture
(Sharma & Leung, 2003; Iliyasu et al., 2014), water utilities (Worthington, 2014),
transport (Markovits-Somogyi, 2011), airports (Merkert et al., 2012; Liebert &
Niemeier, 2013; Iyer & Jain, 2019; Cui & Yu, 2021), seaports (Panayides et al.,
2009; Gonzalez & Trujillo, 2009; Kmac & Kaleibar, 2022), information systems
(Wang et al., 2009), supply chains (Agrell & Hatami-Marbini, 2013; Soheilirad
et al., 2018; Dutta et al., 2022), sustainability (Vorosmarty & Dobos, 2020; Tsaples
& Papathanasiou, 2021; Cikovic et al., 2022), health care (Worthington, 2004;
Zakowska & Godycki-Cwirko, 2020), the wood industry (Salehirad & Sowlati,
2006), electricity distribution (Qassim et al., 2005), and government (Worthington
& Dollery, 2000). These articles contain many novel applications of DEA in their
focal areas.
16 1 Introduction
References
Agrell, P. J., & Hatami-Marbini, A. (2013). Frontier-based performance analysis models for supply
chain management: State of the art and research directions. Computers and Industrial Engi-
neering, 66, 567–583.
Aigner, D. J., & Chu, S. F. (1968). On estimating the industry production function. American
Economic Review, 58, 826–839.
Balezentis, T. (2014). On measures of the agricultural efficiency – A review. Transformations in
Business and Economics, 13(3), 110–131.
Banker, R. D., Charnes, A., & Cooper, W. W. (1984). Some models for estimating technical and
scale efficiencies in data envelopment analysis. Management Science, 30, 1078–1092.
Berger, A. N., & Humphrey, D. B. (1997). Efficiency of financial institutions: International survey
and directions for future research. European Journal of Operational Research, 98, 175–212.
Byrnes, P., Färe, R., & Grosskopf, S. (1984). Measuring productive efficiency: An application to
Illinois strip mines. Management Science, 30, 671–681.
Chachuli, F. S. M., Ludin, N. D., Mat, S., & Sopian, K. (2020). Renewable energy performance
evaluation studies using the data envelopment analysis (DEA): A systematic review. Journal of
Renewable and Sustainable Energy, 12, 062701.
Charnes, A., Cooper, W. W., & Rhodes, E. (1978). Measuring the efficiency of decision making
units. European Journal of Operational Research, 2, 429–444.
Charnes, A., Cooper, W. W., Golany, B., Halek, R., Klopp, G., Schmitz, E., & Thomas, D. (1986).
Two-phase data envelopment analysis approaches to policy evaluation and management of
army recruiting activities: Tradeoffs between joint services and army advertising. Research
Report CCS #532. Center for Cybernetic Studies, University of Texas-Austin.
Cikovic, K. F., Martincevic, I., & Lozic, J. (2022). Application of data envelopment analysis (DEA)
in the selection of sustainable suppliers: A review and a bibliometric analysis. Sustainability,
14, 6672.
Cook, W. D., & Seiford, L. M. (2009). Data envelopment analysis (DEA) – Thirty years on. -
European Journal of Operational Research, 192, 1–17.
Cook, W. D., Liang, L., & Zhu, J. (2010). Measuring performance of two-stage network structures
by DEA: A review and future perspective. Omega, 38, 423–430.
Cui, Q., & Yu, L. T. (2021). A review of data envelopment analysis in airline efficiency: State of the
art and prospects. Journal of Advanced Transportation, Article No. 2931734.
Dutta, P., Jaikumar, B., & Arora, M. S. (2022). Applications of data envelopment analysis in
supplier selection between 2000 and 2020: A literature review. Annals of Operations Research,
315, 1399–1454.
Emrouznejad, A., Parker, B. R., & Tavares, G. (2008). Evaluation of research in efficiency and
productivity: A survey and analysis of the first 30 years of scholarly literature in DEA. Socio-
Economic Planning Sciences, 42, 151–157.
Färe, R., & Grosskopf, S. (2000). Network DEA. Socio-Economic Planning Sciences, 34, 35–49.
Farrell, M. J. (1957). The measurement of productive efficiency. Journal of the Royal Statistical
Society, Series A, 120(Part 3), 253–281.
Farrell, M. J., & Fieldhouse, M. (1962). Estimating efficient production under increasing returns to
scale. Journal of the Royal Statistical Society, Series A, 125(Part 2), 252–267.
Ferguson, C. E., & Gould, J. P. (1986). Microeconomic theory (6th ed.). Irwin.
Fethi, M. D., & Pasiouras, F. (2010). Assessing bank efficiency and performance with operational
research and artificial intelligence techniques: A survey. European Journal of Operational
Research, 204, 189–198.
Førsund, F. R., Lovell, C. A. K., & Schmidt, P. (1980). A survey of frontier production functions
and of their relationship to efficiency measurement. Journal of Econometrics, 13, 5–25.
Gattoufi, S., Oral, M., & Reisman, R. (2004). A data envelopment analysis literature: A bibliogra-
phy updates (1996-2001). Socio-Economic Planning Sciences, 38(2), 159–229.
References 17
Gonzalez, M. M., & Trujillo, L. (2009). Efficiency measurement in the port industry: A survey of
the empirical evidence. Journal of Transport Economics and Policy, 43, 157–192.
Iliyasu, A., Mohamed, Z. A., Ismail, M. M., Abdullah, A. M., Kamarudin, S. M., & Mazuki,
H. (2014). A review of production frontier research in aquaculture (2001-2011). Aquaculture
Economics and Management, 18, 221–247.
Iyer, K. C., & Jain, S. (2019). Performance measurement of airports using data envelopment
analysis: A review of methods and findings. Journal of Air Transport Management, 81, 101707.
Kao, C. (2014). Network data envelopment analysis: A review. European Journal of Operational
Research, 239, 1–16.
Kao, C., & Hwang, S. N. (2008). Efficiency decomposition in two-stage data envelopment analysis:
An application to non-life insurance companies in Taiwan. European Journal of Operational
Research, 185, 418–429.
Kao, C., & Hwang, S. N. (2010). Efficiency measurement for network systems: IT impact on firm
performance. Decision Support Systems, 48, 437–446.
Kmac, E., & Kaleibar, M. M. (2022). A comprehensive review of data envelopment analysis (DEA)
methodology in port efficiency evaluation. Maritime Economics and Logistics. https://doi.org/
10.1057/s41278-022-00239-5
Lampe, H. W., & Hilgers, D. (2015). Trajectories of efficiency measurement: A bibliometric
analysis of DEA and SFA. European Journal of Operational Research, 240, 1–21.
Liebert, V., & Niemeier, H. M. (2013). A survey of empirical research on the productivity and
efficiency measurement of airports. Journal of Transport Economics and Policy, 47, 157–189.
Liu, J. S., Lu, L. Y. Y., Lu, W. M., & Lin, B. J. Y. (2013a). A survey of DEA applications. Omega,
41, 893–902.
Liu, J. S., Lu, L. Y. Y., Lu, W. M., & Lin, B. J. Y. (2013b). Data envelopment analysis 1978-2010:
A citation-based literature survey. Omega, 41, 3–15.
Markovits-Somogyi, R. (2011). Measuring efficiency in transport: The state of the art of applying
data envelopment analysis. Transport, 26, 11–19.
Merkert, R., Odeck, J., Brathen, S., & Pagliari, R. (2012). A review of different benchmarking
methods in the context of regional airports. Transport Reviews, 32, 379–395.
Panayides, P. M., Maxoulis, C. N., Wang, T. F., & Ng, K. Y. A. (2009). A critical analysis of DEA
applications to seaport economic efficiency measurement. Transport Reviews, 29, 183–206.
Panwar, A., Olfati, M., Pant, M., & Snasel, V. (2022). A review on the 40 years of existence of data
envelopment analysis models: Historic development and current trend. Archives of Computa-
tional Methods in Engineering, 29, 5397–5426.
Paradi, J. C., & Zhu, H. Y. (2013). A survey on bank branch efficiency and performance research
with data envelopment analysis. Omega, 41, 61–79.
Qassim, R. Y., Corso, G., Lucena, L. D. S., & Thome, Z. D. (2005). Application of data
envelopment analysis in the performance evaluation of electricity distribution: A review.
International Journal of Business Performance Management, 7, 60–70.
Rhodes, E. (1978). Data Envelopment Analysis and Related Approaches for Measuring the
Efficiency of Decision Making Units with An Application to Program Follow Through in
U.S. Education. Ph.D. Thesis, Carnegie-Mellon University, School of Urban and Public Affairs,
Pittsburgh, PA.
Salehirad, N., & Sowlati, T. (2006). Productivity and efficiency assessment of the wood industry: A
review with a focus on Canada. Forest Products Journal, 56, 11–12.
Seiford, L. M. (1996). Data envelopment analysis: The evolution of the state of the art (1978-1995).
Journal of Productivity Analysis, 7, 99–138.
Sharma, K. R., & Leung, P. S. (2003). A review of production frontier analysis for aquaculture
management. Aquaculture Economics and Management, 15–34.
Shephard, R. W. (1970). Theory of cost and production functions. Princeton University Press.
Soheilirad, S., Govindan, K., Mardani, A., Zavadskas, E. K., Nilashi, M., & Zakuan, N. (2018).
Application of data envelopment analysis models in supply chain management: A systematic
review and meta-analysis. Annals of Operations Research, 271, 915–969.
18 1 Introduction
Song, M., An, Q., Zhang, W., Wang, Z., & Wu, J. (2012). Environmental efficiency evaluation
based on data envelopment analysis: A review. Renewable and Sustainable Energy Reviews, 16,
4465–4469.
Tsaples, G., & Papathanasiou, J. (2021). Data envelopment analysis and the concept of sustain-
ability: A review and analysis of the literature. Renewable and Sustainable Energy Reviews,
138, 110664.
Vorosmarty, G., & Dobos, I. (2020). A literature review of sustainable supplier evaluation in data
envelopment analysis. Journal of Cleaner Production, 264, 121672.
Wang, H. W., Liu, X., & Ding, J. J. (2009). Research on the performance of information systems
with data envelopment analysis. Industrial Engineering and Management, 14, 91–97.
Worthington, A. (2004). Frontier efficiency measurement in health care: A review of empirical
techniques and selected applications. Medical Care Research and Review, 61, 135–170.
Worthington, A. (2014). A review of frontier approaches to efficiency and productivity measure-
ment in urban water utilities. Urban Water Journal, 11, 55–73.
Worthington, A., & Dollery, B. (2000). An empirical survey of frontier efficiency measurement
techniques in local government. Local Government Studies, 26, 23–52.
Xu, T., You, J. X., Li, H., & Shao, L. N. (2020). Energy evaluation based on data envelopment
analysis: A literature review. Energies, 13, 3548.
Zakowska, I., & Godycki-Cwirko, M. (2020). Data envelopment analysis applications in primary
health care: A systematic review. Family Practice, 37, 147–153.
Zhou, P., Ang, B. W., & Poh, K. L. (2008). A survey of data envelopment analysis in energy and
environmental studies. European Journal of Operational Research, 189, 1–18.
Chapter 2
Output–Input Ratio Efficiency Measures
Ever since the industrial revolution, people have been working to use the smallest
effort to produce the largest output, so that resources, including human, are utilized
more efficiently. Manufacturing companies develop standards to help achieve this,
such as the number of items that should be produced with one unit of a certain type of
input, in order to better control the production process and increase productivity.
Similarly, service companies aim to increase the number of customers served by one
employee in a unit of time.
Productivity is generally defined as the amount of output produced by one unit of
input. Theoretically, there is a maximum productivity which can be achieved only
under perfect conditions. The productivity of a production unit divided by the
maximum productivity is the efficiency of this particular unit. In this context,
efficiency is always less than or equal to unity.
Consider a production activity that applies multiple inputs to produce multiple
outputs. Let Xij, i = 1, . . ., m be the quantity of input i employed by unit j in a period
of time, and Yrj, r = 1, . . ., s be the quantity of output r produced in the same period.
Productivity is thus ratio of the aggregate output to the aggregate input, expressed as:
s
r = 1 ur Y rj
Pj = m ð2:1Þ
i = 1 vi X ij
where Pj is the productivity of unit j, and ur and vi are the conversion factors
(or weights) of output r and input i, respectively (Bitran & Chang, 1984). The key
point in this measure is the determination of the weights, so that all the inputs and
outputs are restated in their common denominations. Prices are commonly used in
such calculations, such that the productivity indicates the amount of money that can
be generated from each dollar consumed. This formula is quite simple and easy to
calculate. However, the problem is that in many cases some inputs and outputs do
not have market values, which makes the aggregation of the inputs and outputs
difficult. For example, in calculating the productivity of a forest, it is difficult to

https://doi.org/10.1007/978-3-031-27593-7_2
20 2 Output–Input Ratio Efficiency Measures
determine the monetary values of soil conservation, carbon dioxide absorption, and
wildlife habitation. For this reason, the measurement of productivity was for a long
time limited to the ratio of one output to one input when there is the problem of
commensurability.
2.1 CCR Model
Charnes et al. (1978) proposed a fractional programming model, commonly referred

to as the CCR model, in which the problem of non-commensurability was solved.
The idea is to allow the focal production unit (generalized as the decision making
unit, DMU, in their study) to select the most favorable weights (or multipliers) ur and
vi to calculate the productivity ratio. The only restriction is that the productivity
ratios of all DMUs calculated from the multipliers selected by this DMU must be less
than or equal to one. Since this ratio is between zero and one, and it can be shown
that the ratio is equal to the actual output of a DMU to the maximum output that can
be produced with the same amount of input of this DMU, it is a measure of
efficiency. It is also a relative measure when the maximum output is obtained from
a sample.
2.1.1 Input Model
The CCR model for measuring the relative efficiency of a DMU indexed by 0 is:
s
r = 1 ur Y r0
E 0 = max: m
i = 1 vi X i0
s
r = 1 ur Y rj
ð2:2Þ
s:t: m ≤ 1, j = 1, . . . , n
i = 1 vi X ij
ur , vi ≥ ε, r = 1, . . . , s, i = 1, . . . , m
Note that the multipliers ur and vi are required to be greater than a small positive
number ε, to avoid some unfavorable factors being ignored by assigning zero to the
corresponding multipliers (Charnes et al., 1979). This small number ε is called a
non-Archimedean number (Charnes & Cooper, 1984). If E0 = 1, then this DMU is in
a state of Pareto optimality, also called Pareto efficiency (Koopmans, 1951). Orig-
inally, Pareto efficiency referred to a state that augments the value of one variable
necessarily reduces the value of another. Koopmans (1951) extended it to productive
efficiency to refer to a state that an improvement in any factor, i.e., an increase in an
output or a decrease in an input, requires a deterioration of at least one other factor,
2.1 CCR Model 21
i.e., a decrease in at least one output or an increase in at least one input. It thus is also
called Pareto-Koopmans efficiency (Charnes & Cooper, 1961). If the lower bound ε
is removed, and one still has E0 = 1, then this DMU is only weakly efficient
(Charnes et al., 1986, 1991), because in this case one can reduce the amount of
some input or increase the amount of some output and still have E0 = 1. In contrast
to the weakly efficient condition, the normal case of E0 = 1, with the lower bound ε
imposed upon the multipliers, is called strongly efficient.
Model (2.2) is in ratio form. It is a linear fractional program, which, based on the
ideas set out in Charnes and Cooper (1962), can be transformed into the following
linear model:
s
E 0 = max: ur Y r0
r=1
m
s:t: vi X i0 = 1
i=1 ð2:3Þ
s m
ur Y rj - vi X ij ≤ 0, j = 1, . . . , n
r=1 i=1
ur , vi ≥ ε, r = 1, . . . , s, i = 1, . . . , m
This form of the model is called a multiplier form. When there is only one input
and one output this model uses a straight line, passing through the origin and
superimposing upon all DMUs, as the production frontier. Since the production
frontier passes through the origin, this implies that a proportional change in the input
leads to the same proportional change in the output. One thus has a situation of
constant returns to scale.
Model (2.3) has a dual, which can be formulated as:
m s
E 0 = min: θ - ε si- þ sþ
r
i=1 r=1
n
s:t: λj X ij þ si- = θX i0 , i = 1, . . . , m
j=1
n ð2:4Þ
λj Y rj - sþ
r = Y r0 , r = 1, . . . , s
j=1
λj , si- , sþ
r ≥ 0, j = 1, . . . , n, i = 1, . . . , m, r = 1, . . . , s
θ unrestricted in sign:
Since si- , sþ
r ≥ 0, the first two sets of constraints imply
n
j = 1 λj X ij ≤ θX i0 and
n
j = 1 λj Y rj
≥ Y r0 , which indicate that all observations have a larger amount of inputs
and smaller amount of outputs than the point ( nj = 1 λj X ij , nj = 1 λj Y rj ) on the
production frontier. In other words, the observations are enveloped by the
Table 2.1 CCR efficiency of DMU Input X Output Y Efficiency

an example with one input and
A 3 2 2/3
one output
B 4 4 1
C 5 3 0.6
D 6 6 1
E 10 8 0.8
Fig. 2.1 CCR efficiency Y

with one input and one
output
10 R
8 E
D
6
YV V
B
4
YC U
C
2 A
XU XC
O 2 4 6 8 10 X
production frontier. Model (2.4) is thus called an envelopment form of the model.
Moreover, the second constraint set indicates that the DMU fixes its outputs at the
current level of Yr0 (or an adjusted amount of Y r0 þ sþ r , to be exact) to look for the
reduction ratio θ that the amount of inputs can be reduced. Model (2.4), or Model
(2.2), is thus an input model. Another point to be noted is that although mathemat-
ically θ is not restricted to be positive, it will always be a positive number, less than
or equal to one, due to the basic properties of the problem (as the smallest reduction
ratio is zero).
Consider five DMUs, A, B, C, D, and E, each applying input X to produce output
Y, with the data shown in Table 2.1, and as depicted in Fig. 2.1. Ray OR, lying above
all observations with the smallest slope, is the production frontier constructed from
these DMUs. The efficiency of each DMU is the ratio of the minimum input needed
to produce the same amount of output of this DMU (on the production frontier) to the
actual amount of input used by this DMU. For example, the efficiency of DMU C is
XU/XC = 0.6. By applying Model (2.2), the efficiency of all DMUs can be calculated,
with the results shown in the last column of Table 2.1.
2.1 CCR Model 23
Table 2.2 CCR efficiency of DMU Input X1 Input X2 Output Y Efficiency

an example with two inputs
A 1 4 1 1
and one output
B 2 2 1 1
C 4 1 1 1
D 4 3 1 0.6
E 6 1 1 1-2ε
Fig. 2.2 CCR efficiency X2

measured from the isoquant
4 A
2
B
G E
C
O 2 4 6 X1
Consider another case, where five DMUs, A, B, C, D, and E, applying different

combinations of two inputs X1 and X2 to produce one unit of output Y, with the data
shown in Table 2.2. Fig. 2.2 shows the isoquant of Y = 1, constructed from these
DMUs, where DMUs A, B, and C are efficient, as they lie on the isoquant. DMU
D has an efficiency score of OG/OD = 0.6. DMU E lies on the part of the isoquant
extended horizontally from DMU C, which indicates that it is weakly efficient. This
is a situation in which by assigning zero to multiplier v1 in Model (2.2), one obtains
EE = 1. As a matter of fact, DMU E is dominated by DMU C, as it consumes two
more units of X1 than DMU C to produce the same amount of Y. The associated
optimal solution is u = 1 - 2ε, v1 = ε, and v2 = 1 - 6ε.
2.1.2 Output Model
Efficiency can also be measured from the output side. The CCR efficiency in this
case is represented in the reciprocal form of 1/E0. The full output model in ratio
form is:
m
vi X i0
1 i=1
= min: s
E0
ur Y r0
r=1
m ð2:5Þ
vi X ij
i=1
s:t: s ≥ 1, j = 1, . . . , n
ur Y rj
r=1
ur , vi ≥ ε, r = 1, . . . , s, i = 1, . . . , m:
This model is exactly the same as Model (2.2), except that the objective function
is represented in reciprocal form. Its linear transformation, in multiplier form, is:
m
1
= min: vi X i0
E0 i=1
s
s:t: ur Y r0 = 1
r=1
ð2:6Þ
m s
vi X ij - ur Y rj ≥ 0, j = 1, . . . , n
i=1 r=1
ur , vi ≥ ε, r = 1, . . . , s, i = 1, . . . , m,
and the corresponding envelopment form of the model, which is the dual of Model
(2.6), is:
m s
1
= max: φ þ ε si- þ sþ
r
E0 i=1 r=1
n
s:t: λj X ij þ si- = X i0 , i = 1, . . . , m
j=1
ð2:7Þ
n
λj Y rj - sþ
r = φY r0 , r = 1, . . . , s
j=1
λj , si- , sþ
r ≥ 0, j = 1, . . . , n, i = 1, . . . , m, r = 1, . . . , s
φ unrestricted in sign:
The constraints of this model indicate that this model fixes the inputs at the
current level of Xi0 (or the adjusted amount of X i0 - si- , to be exact) and looks for
the largest extent φ that the outputs can be expanded. For this reason, this model is an
output model.
Referring to the DMUs in Fig. 2.1, the output model measures the efficiency of
DMU C by using point V on the production frontier as the benchmark, which applies
the same amount of input as DMU C does to produce the largest possible amount of
output. Geometrically, the inverse of the efficiency, as calculated from Model (2.6),
2.1 CCR Model 25
Table 2.3 CCR efficiency of DMU Input X Output Y1 Output Y2 Efficiency

an example with one input and
A 1 2 3 2/3
two outputs
B 1 2 5 1
C 1 4 1 1/(1.25+0.75ε)
D 1 4 4 1
E 1 5 2 1
Fig. 2.3 CCR efficiency Y2

measured from the product
transformation curve
6
B
G
4 D
2 E
H
C
O 2 4 6 Y1
is the ratio of YV to YC. Under constant returns to scale, which is reflected by the
linear production frontier passing through the origin, the efficiency measured from
the input model, XU/XC, is the same as that measured from the output model, YC/YV,
although the benchmarks selected by the input and output models are different.
To discuss efficiency measurement from the output side in higher dimensions,
consider five DMUs, A, B, C, D, and E, each applying one unit of input X to produce
different amounts of two outputs, Y1 and Y2, with the data shown in Table 2.3. The
product transformation curve constructed from these DMUs is depicted in Fig. 2.3,
where DMUs B, D, and E, lying on the product transformation curve, are efficient.
This figure shows that DMU A uses point G, on the product transformation curve, as
the benchmark to measure its efficiency, with EA = OA/OG = 2/3. DMU C uses
point H, on the segment of the product transformation curve extended almost
vertically from DMU E, to measure efficiency, to get an efficiency score of OC/
OH = 1/(1.25 + 0.75ε). The optimal solution is
v = 1:25 þ 0:75ε, u1 = 0:25ð1 - εÞ, u2 = ε. Note that 0.75ε in the denominator
is caused by the very small scale of the vertical line tilting to the right from point E,
as required by u2 ≥ ε.
When a set of DMUs uses the same amount of input to produce different amounts
of output, it is obvious that the one with the largest output has a relative efficiency
equal to one, and all others have a relative efficiency equal to the ratio of their
amount of output to the largest amount. The efficiency measured from the CCR
model satisfies this definition. Consider a situation of n DMUs, all applying the same
amount of m inputs, Xik = Xij, j = 1, . . ., n, i = 1, . . ., m, to produce different
amounts of one output Y. From Fig. 2.1 it is clear that the DMU with the largest
amount of output, Ymax, must lie on the production frontier. Let
(u , vi , i = 1, . . . , m) be the optimal solution obtained from Model (2.2) in calcu-
lating the efficiency of a DMU. For the DMU with the largest amount of output Ymax,
one has u Y max = m
i = 1 vi X i max =
m
i = 1 vi X io : The efficiency of the DMU being

evaluated can be expressed as: E0 = u Y 0 = m i = 1 vi X i0 = u Y 0 =u Y max = Y 0 =Y max ,
as expected.
Consider another situation, where all DMUs apply different amounts of one input
X to produce the same amount of s outputs, i.e., Yr0 = Yrj, j = 1, . . ., n, r = 1, . . .,
s. The efficiency of the DMU being evaluated, from the input point of view, is the
ratio of the minimum input level, Xmin, divided by the actual amount of input
consumed: Xmin/X0. Let (ur , r = 1, . . ., s, v) be the optimal solution obtained
from Model (2.2) in calculating the efficiency of a DMU. By the same token, one
has sr = 1 ur Y r0 = sr = 1 ur Y r min = v X min : The efficiency of the DMU being eval-
uated can then be expressed as: E 0 = sr = 1 ur Y r0 =v X 0 = v X min =v X 0 = X min =X 0 ,
as expected. This verifies that the efficiency measured from the CCR model follows
the conventional definition of relative efficiency.
2.2 BCC Model
The CCR model assumes constant returns to scale, in that the output increases in the
same proportion as the input. For the case of a single input and single output, the
production frontier is a straight line passing through the origin. In production
economics, due to the effect of fixed inputs, returns to scale usually increase in the
early stage of production, where the amount of variable input is relatively small. As
the amount of variable input increases, returns to scale diminish to constant, and
finally become decreasing. Taking this phenomenon into consideration, Banker et al.
(1984) extended the CCR model to allow for variable returns to scale, referred to as
the BCC model. Conceptually, they allow the production frontier to move away
from the origin by introducing a constant in aggregating either the inputs or outputs.
The constant plays the role of the intercept in the linear production frontier. This
model also has two forms: input and output.
2.2 BCC Model 27
2.2.1 Input Model
The idea of the output–input ratio efficiency measure is to aggregate the outputs into
a virtual output and the inputs into a virtual input and to take their ratio to be the
measure of efficiency. In the input model the outputs are treated as the explanatory
variables to calculate the expected (minimum) virtual input. The ratio of the mini-
mum virtual input to the actual virtual input, aggregated from the multiple inputs, is
the input efficiency. The model developed by Banker et al. (1984) to measure the
efficiency from the input side is:
s
ur Y r0 - u0
r=1
E0 = max: m
vi X i0
i=1
s
ur Y rj - u0 ð2:8Þ
r=1
s:t: m ≤ 1, j = 1, . . . , n
vi X ij
i=1
ur , vi ≥ ε, r = 1, . . . , s, i = 1, . . . , m
u0 unrestricted in sign:
The difference between this model and Model (2.2), the one under constant
returns to scale, is the inclusion of the intercept u0.
The linear fractional objective function in Model (2.8) can be linearized by
assigning the denominator to one and leaving the numerator as the objective
function. This is because Model (2.8) has multiple solutions, in that if (u, v) is
an optimal solution, then so is (cu, cv), for c > 0. Assigning the denominator to one
to reduce one degree of freedom will thus not alter the optimal objective value, E0,
although the optimal solution (u, v) may be different. The linear fractional con-
straints are easily linearized by multiplying both sides by the denominator to obtain
the following linear programming model:
s
E 0 = max: ur Y r0 - u0
r=1
m
s:t: vi X i0 = 1
i=1
s m ð2:9Þ
ur Y rj - u0 - vi X ij ≤ 0, j = 1, . . . , n
r=1 i=1
ur , vi ≥ ε, r = 1, . . . , s, i = 1, . . . , m
This linear model can also be obtained by applying the idea in Charnes and
Cooper (1962) for transforming a linear fractional program into a linear program.
Table 2.4 Technical and scale efficiencies of an example with one input and one output
BCC Input Eff. BCC Output Eff.
DMU Input X Output Y CCR Eff. Technical Scale Technical Scale
A 3 2 2/3 1 2/3 1 2/3
B 4 4 1 1 1 1 1
C 5 3 0.6 0.7 6/7 0.6 1
D 6 6 1 1 1 1 1
E 10 8 0.8 1 0.8 1 0.8
m s
E0 = min: θ - εð si- þ sþ
r Þ
i=1 r=1
n
j=1
n
λj Y rj - sþ
r = Y r0 , r = 1, . . . , s ð2:10Þ
j=1
n
λj = 1
j=1
λj , si- , sþ
r ≥ 0, j = 1, . . . , n, i = 1, . . . , m, r = 1, . . . , s:
This model is almost the same as Model (2.4), the one under the assumption of
constant returns to scale, except that a convexity constraint of nj = 1 λj = 1 is added.
To see the difference between the BCC model and the CCR model in measuring
efficiencies from the input side, consider the example in Table 2.1, where five DMUs
apply one input X to produce one output Y. The first four columns of Table 2.4 are
copied from Table 2.1 for easy comparison, and the positions of the five DMUs in
the X-Y plane are redrawn in Fig. 2.4 for detailed explanation. By applying Model
(2.8), one obtains the efficiencies of the five DMUs, as shown in the fifth column of
Table 2.4, where DMUs A, B, D, and E are efficient, and only DMU C is inefficient.
This implies that the production frontier constructed from these DMUs by this model
is composed of the piecewise line segments ABDE, as shown in Fig. 2.4. Since this
problem has only one input, we can divide the multipliers in the numerator by that in
the denominator to obtain the following linear program and have a clearer geometric
interpretation:
2.2 BCC Model 29
10 R
8
E
6
D
B V
4
v0 U
YW (YC) W C
2 A
– u0 – u0 XU XW XC
–6 –4 –2 O 2 4 6 8 10 X
–2
–4 v0
Fig. 2.4 Efficiency measurement under variable returns to scale
E 0 = max: ðu1 Y 0 - u0 Þ=X 0

s:t: ðu1 Y j - u0 Þ=X j ≤ 1, j = A, B, C, D, E ð2:11Þ
u1 ≥ ε, u0 unrestricted in sign:
The multiplier u1 in this case is the slope of the line segment corresponding to the
DMU being evaluated by considering Y as the horizontal axis, and u0 is the intercept
of the line segment extending to the X-axis. For example, in calculating the effi-
ciency of DMU A, one obtains u1 = 0:5 and - u0 = 2, corresponding to the line
segment AB. From Fig. 2.4 it is noted that this is not the unique solution, and
u1 , - u0 = ðε, 3 - 2εÞ, which represents a vertical line extended from DMU A,
is obviously another one. As a matter of fact, any line passing through DMU A with a
slope u1 between ε and 0.5, and the corresponding intercept of -u0 = 3 - 2u1, can
be the frontier. By the same token, when Model (2.11) is applied to measure the
efficiency of DMUs B, D, and E, one also obtains alternative solutions. Based on
Fig. 2.4, the alternative solutions for DMU B are 0:5 ≤ u1 ≤ 1, with - u0 = 4 - 4u1 ,
for DMU D they are 1 ≤ u1 ≤ 2, with - u0 = 6 - 6u1 , and for DMU E they are
2 ≤ u1 < 1, , with - u0 = 10 - 8u1 :
In calculating the efficiency of DMU C by applying Model (2.11), the optimal
solution is unique, with u1 , - u0 = ð0:5, 2Þ, in that the line segment AB is the
frontier, and point W is the benchmark, that DMU C uses when measuring efficiency.
The efficiency is XW/XC = 0.7. Recall that the frontier used to calculate efficiency
under constant returns to scale is the ray OR, with the CCR efficiency of 0.6. From
Fig. 2.4 it is clear that for the whole production frontier the region under the line
segment BD (in terms of the output) has the most productive scale, in that the
average input consumed per unit of output is the smallest. For scales (in terms of
output) smaller than that of DMU B or greater than that of DMU D, the average
amount of input used to produce one unit of output is larger, and the larger amount of
input needed in these two regions is due to inadequate scales. The extra amount of
input required relative to the minimum amount is the inefficiency due to improper
scale. For example, point W, with a BCC efficiency one, is technically efficient;
however, it is not efficient from the viewpoint of scale, because under the most
productive scale an input level of XU is enough to produce the output level YW. XU/
XW is thus the scale efficiency of W. Since DMU C has the same output level of W, it
has the same scale efficiency as W has, XU/XW. Different from W, DMU C is
technically inefficient, because it requires XC units of input, rather than XW, to
produce the same amount of output YC, and XC/XW is its technical efficiency. By
generalizing this idea to multiple-input and multiple-output cases, one has the BCC
efficiency as the technical efficiency, and the ratio of the CCR efficiency to the BCC
efficiency as the scale efficiency for the DMU being evaluated. Column six of
Table 2.4 shows the scale efficiencies of the five DMUs.
In higher dimensions the optimal solution (u, v) for a technically efficient DMU
has sr = 1 ur Y r - u0 - m
i = 1 vi X i = 0, where Xi, Yr are now variables, which is a
supporting hyperplane. Banker et al. (1984) showed that when the solution is unique,
negative, zero, and positive values of u0 indicate that the associated DMU is in the
regions of increasing, constant, and decreasing returns to scale, respectively. This is
also seen in the graphical interpretation shown in Fig. 2.4. Banker and Thrall (1992)
presented a proof for this when the focal DMU is technically efficient, and Banker
et al. (1996) later removed this condition.
2.2.2 Output Model
In contrast to the input model, where the minimum amount of inputs needed to
produce the specified output levels is obtained to measure efficiency, the output
model looks for the maximum amount of outputs that can be produced from the
given amount of inputs to measure efficiency. Based on this concept, the output BCC
model for measuring the efficiency of a DMU has the following form:
2.2 BCC Model 31
m
vi X i0 þ v0
1 i=1
= min: s
E0
ur Y r0
r=1
m
vi X ij þ v0 ð2:12Þ
i=1
s:t: s ≥ 1, j = 1, . . . , n
ur Y rj
r=1
ur , vi ≥ ε, r = 1, . . . , s, i = 1, . . . , m
v0 unrestricted in sign:
Similar to the input model, the numerator of the fractional in Model (2.12)
calculates the maximal virtual output that can be produced from the actual amount
of inputs for the DMU being evaluated, and the denominator calculates the virtual
output aggregated from the actual amount of outputs. The ratio is then the reciprocal
of the efficiency, from the output point of view.
The equivalent multiplier form to the ratio form (2.12) of the model is:
m
1
= min: vi X i0 þ v0
E0 i=1
s
s:t: ur Y r0 = 1
r=1
ð2:13Þ
m s
vi X ij þ v0 - ur Y rj ≥ 0, j = 1, . . . , n
i=1 r=1
ur , vi ≥ ε, r = 1, . . . , s, i = 1, . . . , m
v0 unrestricted in sign
and the corresponding envelopment form of the model, which is the dual of Model
(2.13), is:
m s
1
= max: φ þ ε si- þ sþ
r
E0 i=1 r=1
n
s:t: λj X ij þ si- = X i0 , i = 1, . . . , m,
j=1
n
λj Y rj - sþ
r = φY r0 , r = 1, . . . , s: ð2:14Þ
j=1
n
λj = 1
j=1
λj , si- , sþ
r ≥ 0, j = 1, . . . , n, i = 1, . . . , m, r = 1, . . . , s:
This envelopment form of the model also has an assumption of variable returns to
scale, and the difference of this from that under constant returns to scale, i.e., Model
Table 2.5 Summary of the BCC models in different forms

Input model Output model
Ratio form
s m
r = 1 ur Y r0 - u0 1 i = 1 vi X i0 þ v0
E 0 = max: m = min: s
i = 1 vi X i0 E0 r = 1 ur Y r0
s m
r = 1 ur Y rj - u0 i = 1 vi X ij þ v0
s:t: m ≤ 1, j = 1, . . . , n s:t: s ≥ 1, j = 1, . . . , n
i = 1 vi X ij r = 1 ur Y rj
ur , vi ≥ ε, r = 1, . . . , s, i = 1, . . . , m ur , vi ≥ ε, r = 1, . . . , s, i = 1, . . . , m
u0 unrestricted in sign: v0 unrestricted in sign:
Multiplier form
s
E 0 = max: r = 1 ur Y r0 - u0
1
= min:
m
v X þ v0
m E0 i = 1 i i0
s:t: i = 1 vi X i0 = 1 s
s m s:t: r = 1 Y r0 = 1
u
r = 1 ur Y rj - u0 - ≤ 0, r
i = 1 vi X ij m s
j = 1, . . . , n i = 1 vi X ij þ v0 - r = 1 ur Y rj ≥ 0,
j = 1, . . . , n
ur , vi ≥ ε, r = 1, . . . , s, i = 1, . . . , m
ur , vi ≥ ε, r = 1, . . . , s, i = 1, . . . , m
Envelopment form
- þ
E 0 = min: θ - ε m i = 1 si þ
s
r = 1 sr
1
= max:φ þ ε
m
s- þ
s
sþ
n - E0 i=1 i r=1 r
s:t: j=1 λj X ij þ s i = θX i0 , i = 1, . . . , m n -
n þ
s:t: j = 1 λj X ij þ si = X i0 , i = 1, . . . , m
j = 1 λj Y rj - sr = Y r0 , r = 1, . . . , s n þ
n j = 1 λj Y rj - sr = φY r0 , r = 1, . . . , s
j = 1 λj = 1, n
λ = 1,
λj , si- , sþ r ≥ 0, j = 1, . . . , n, j=1 j
λj , si- , sþ r ≥ 0, j = 1, . . . , n,
i = 1, . . . , m, r = 1, . . . , s
i = 1, . . . , m, r = 1, . . . , s
(2.7), is the convexity constraint of nj = 1 λj = 1. Table 2.5 summarizes the BCC

input and output models in ratio, multiplier, and envelopment forms.
Model (2.12) can also be used to discuss the region of returns to scale, where the
DMU being evaluated is located. Consider the example in Fig. 2.4. The production
frontier constructed from the five DMUs by Model (2.12) is the same set of line
segments ABDE as that constructed by Model (2.8). Similar to the discussion in the
input model of Sect. 2.2.1, where the only multiplier v in the denominator can be
absorbed by the multipliers u1 and u0 in the numerator, Model (2.12) can be
simplified as follows by merging the only multiplier u in the denominator into v1
and v0 in the numerator:
2.2 BCC Model 33
1
= min:ðv1 X k þ v0 Þ=Y 0
E0
ð2:15Þ
s:t: v1 X j þ v0 =Y j ≥ 1, j = A, B, C, D, E
v1 ≥ ε, v0 unrestricted in sign:
where v1 is the slope, and v0 is the intercept, of the linear frontier that the DMU being
evaluated uses to measure efficiency. For example, to measure the efficiency of
DMU A, one solution is v1 , v0 = ð2, - 4Þ, which corresponds to the line segment
AB. In measuring the efficiencies of DMUs B and D, one solution is (v1 , v0 Þ = ð1, 0Þ,
which corresponds to the line segment BD. Finally, in calculating the efficiency of
DMU E, one solution is v1 , v0 = ð0:5, 3Þ: As discussed in the input model, for the
whole production frontier, the region of constant returns to scale, i.e., line segment
BD, with v0 = 0, has the largest average return (amount of output per unit input), and
the marginal return (the additional amount of output produced by an additional unit
of input) is equal to the average return. In the region of smaller scales (in terms of
input), with v0 < 0, the average return is smaller than that of the region of constant
returns to scale, although the marginal return of the former is larger than that of the
latter. In contrast, in the region of larger scales, with v0 > 0, both the average and
marginal returns are smaller than those of the region of constant returns to scale. The
intercept v0 serves as an indicator of the type of returns to scale of the DMU.
DMU A lies on the production frontier, and it is thus technically efficient.
However, if it is measured by the CCR model, then it is not efficient, with a CCR
efficiency of 2/3, based on the benchmark U. This inefficiency is obviously due to an
inadequate scale. The BCC model (2.12) thus measures the technical efficiency,
while the CCR model (2.5) measures the overall efficiency, and the ratio of the CCR
efficiency to the BCC efficiency is the scale efficiency. It is worth noting that DMU
C has a scale efficiency of one, although it is technically inefficient. The second-to-
last column of Table 2.4 shows the technical efficiencies of the five DMUs from the
output side, and the last column shows their scale efficiencies. The BCC efficiencies
from the input and output sides, including technical and scale, are not necessarily
the same.
The graphical interpretation in this section shows that when the number of DMUs
is relatively small, most of them appear on the production frontier, with a perfect
efficiency score of one, which apparently overstates their efficiency. This raises the
question of how many DMUs are needed to construct an empirical frontier which
does not deviate from the true one by too much, in order to obtain meaningful
efficiency measures. A rule of thumb is to have at least three times the total number
of inputs and outputs, that is n ≥ 3(m + s) (Banker et al., 1989). Time series data can
be used for cases in which the number of all possible DMUs does not satisfy this
rule, by considering the same DMU at different time periods as different DMUs, thus
increasing the number of DMUs in the calculation, referred to as window analysis in
Charnes et al. (1985).
2.3 Restrictions on Multipliers
One issue that was widely discussed in the early development of the DEA approach
was the value used for the small non-Archimedean number ε. If this is too small, then
it will be ignored in computer calculations due to rounding. If, on the other hand, it is
not small enough, then a Pareto efficiency DMU may become inefficient. As to what
value should be assigned to ε, Lewin and Morey (1981) recommended 10-6.
However, since different units of measurement for the input and output factors
will affect the function of ε, e.g., centimeters versus kilometers, it is inappropriate
to assign the same value to ε for factors of different scales. Charnes and Cooper
(1984) thus suggested using ε = 10-5 when efficiency is expressed as a percentage
(e.g., using E0 = 100 rather than 1.0), and input and output entries are kept in the
range of 1 to 100. There are also other suggestions for setting the value of ε (Färe &
Hunsaker, 1986).
In many cases there is prior information regarding the importance of the factors
that requires the corresponding multipliers to lie in specific ranges in the form of
(Dyson & Thanassoulis, 1988):
LIi ≤ vi ≤ U Ii , i = 1, . . . , m
ð2:16Þ
LO
r ≤ ur ≤ U O
r , r = 1, . . . , s
Consider a case of n DMUs, where each DMU applies different amounts of two
inputs X1 and X2 to produce one unit of one output Y. The CCR input model (2.2),
with restrictions on the range of the multipliers included, can be expressed as:
E0 = max: 1=ðv1 X 10 þ v2 X 20 Þ
s:t: v1 X 1j þ v2 X 2j ≥ 1, j = 1, . . . , n ð2:17Þ
LIi ≤ vi ≤ U Ii , i = 1, 2
where (X1j, X2j) is the input observation of a DMU. The lower bound constraint
LIi ≤ vi , i = 1, 2, can be expressed as v1 1=LI1 þ v2 ð0Þ ≥ 1 and v1 ð0Þ þ
v2 1=LI2 ≥ 1: These two constraints imply that two more DMUs, with observations
(1/ LI1 , 0) and (0, 1/LI2 ), are included to construct the frontier.
To handle the upper bound of vi ≤ U Ii , we substitute it into the constraint of
v1X1j + v2X2j ≥ 1 to obtain U I1 X 1j þ v2 X 2j ≥ 1 and v1 X 1j þ U I2 X 2j ≥ 1, or v2 ≥
ð1 - U I1 X 1j Þ=X 2j and v1 ≥ 1 - U I2 X 2j =X 1j . These lower bounds on vi imply the
addition of two sets of n DMUs, with observations 0, X 2j = 1 - U I1 X 1j and
X 1j = 1 - U I2 X 2j , 0 , j = 1, . . ., n, to the original DMUs to construct the frontier
together.
As more DMUs are included, the constructed frontier will be raised higher in the
X-Y plane (or expanded toward the origin in the input space or expanded outwards in
the output space). The efficiency of every DMU will thus either remain the same or
decrease. For the data contained in Table 2.2, suppose the restrictions of
2.3 Restrictions on Multipliers 35
Fig. 2.5 Frontiers adjusted X2

by absolute bounds on
multipliers
20/3 T S
G
B
H C E
S’
T’
O 5 X1
0.2 ≤ v1 ≤ 0.4 and 0.1 ≤ v2 ≤ 0.5 are imposed. The new DMUs generated by the
lower bounds are (5, 0) and (0, 10). The two upper bounds are able to generate ten
new DMUs; however, only three are feasible, (8, 0), (0, 20/3), and (0, 10). Referring
to Fig. 2.2, which is redrawn as Fig. 2.5, the line segments SABCS′ are the frontier
constructed from the original five DMUs, and TBT′ are those constructed from the
new set of DMUs. Based on this new frontier, the original efficient DMU C and the
weak efficient DMU E become inefficient, the efficiency of DMU D decreases from
OG/OD to OH/OD, whereas the original efficient DMUs A and B are still efficient.
The range in Expression (2.16) for each multiplier is in absolute scale, which has
different effects for measures of different scales, and may obtain misleading results
(Podinovski, 1999). To eliminate this undesirable effect, Thompson et al. (1986)
proposed the concept of an assurance region, with the following form, to restrict the
range of the multipliers:
vi
LIi ≤ ≤ U Ii , i = 2, . . . , m
v1
ur ð2:18Þ
r ≤ u ≤ Ur , r = 2, . . . , s
O
LO
1
In this form the importance of each input/output factor is expressed in relation to

that of the first one. The absolute bounds in (2.16) are special cases of the relative
bounds, with v1 = u1 = 1.
To see how these constraints affect the frontier, consider the example in
Table 2.3, where all five DMUs apply one unit of input X to produce different
amounts of two outputs Y1 and Y2. The constraints of the CCR output model (2.5) for
this problem are u1Y1j + u2Y2j ≤ 1, j = A, . . ., E. Figure 2.6 is redrawn from Fig. 2.3,
in which the line segments SBDES′ are the frontier constructed from the original five
Fig. 2.6 Frontiers adjusted

Y2
by relative bounds on
multipliers 8 T
Slope= –1
B
S
G D
C H
Slope=–5
S’ T’
O 5.4 Y1
DMUs. The frontier line segments BD and DE can be expressed by the equation
u1y1 + u2y2 = 1, with a slope of -u1/u2 equal to -0.5 and -2, respectively. Note that
here y1 and y2 are variables. Suppose an assurance region of 1 ≤ u1/u2 ≤ 5 is
imposed. This implies that two frontiers with slopes of -1 and -5 are added.
These two frontiers correspond to line segments TD and ET′ in Fig. 2.6, and the
new frontier becomes the line segments TDET′. Under this new frontier, the origi-
nally efficient DMU B becomes inefficient, the inefficient DMUs A and C have lower
efficiency scores, and the efficient DMUs D and E are still efficient.
The assurance region for restricting the relative range of either input or output
multipliers can be linked to be more general (Thompson et al., 1990), based on the
concept of a cone ratio (Charnes et al., 1989). The most general linear form of
restrictions on multipliers is: α1u1 + . . . + αsus + β1v1 + . . . + βmvm ≥ 0, and Wong
and Beasley (1990) had an application for this. Tracy and Chen (2005) introduced a
formulation which provides generalized treatment for weight restrictions.
2.4 Ranking
An issue closely related to restrictions on multipliers is ranking. The DEA technique

identifies efficient DMUs, and there is usually more than one DMU that is efficient.
All efficient DMUs have a perfect efficiency score of one, which makes ranking of
them difficult. Imposing tighter ranges for the multipliers, as discussed in the
2.4 Ranking 37
preceding section, may help discriminate the efficient DMUs. In addition to weight
restrictions, there are many other approaches to ranking.
DMUs with higher efficiency scores are usually considered more efficient, and
thus have higher ranks. However, some scholars believe that DMUs using different
frontier facets to measure efficiency are not comparable, and only those using the
same frontier facet can be compared. The same frontier facet means the same value
of multipliers for calculating efficiency scores. Based on this idea, some studies use
the same weight for all DMUs to calculate efficiency. This is the most stringent case
of the assurance region, in which there is only one set of multipliers that can be
selected in the region.
The first article to propose this idea was Roll et al. (1991), using the multipliers
that yield the largest average efficiency score from all DMUs to calculate the
efficiency of every DMU. The model under constant returns to scale is:
n
1
max: Ej
n j=1
s
ur Y rj ð2:19Þ
r=1
s:t: E j = m ≤ 1, j = 1, . . . , n
vi X ij
i=1
ur , vi ≥ ε, r = 1, . . . , s, i = 1, . . . , m
At optimality, the efficiency of DMU j is calculated as

E j = sr = 1 ur Y rj = m
i = 1 vi X ij . The efficiency scores thus measured then have a
common basis for ranking. The common-weight frontier is a hyperplane that super-
imposes upon all DMUs, and all DMUs use this frontier to calculate efficiency.
Kao and Hung (2005) proposed the idea of using the compromise solution to
determine the set of common weights that minimizes the total difference between the
ideal efficiency (calculated from the conventional CCR or BCC models) and the
actual efficiency (calculated from the common weight) of all DMUs to determine the
multipliers. The model is:
s p
n ur Y rj
min: E j - r=1
m
j=1 vi X ij
i=1
s
ur Y rj ð2:20Þ
r=1
s:t: m ≤ 1, j = 1, . . . , n
vi X ij
i=1
ur , vi ≥ ε, r = 1, . . . , s, i = 1, . . . , m:
where Ej in this case is the efficiency of DMU j calculated from the CCR model, and
p > 1 is the distance parameter. When p = 1 this model is equivalent to Model
(2.19), and it can thus be considered as an extension of the idea of maximizing the
average efficiency of all DMUs. Kao and Hung (2005) recommended using p = 2, as
this value produces a result of minimum variance.
Another idea related to a common weight is cross efficiency (Doyle & Green,
1994), and this approach uses the multipliers selected by DMU j to calculate the
efficiency of all other DMUs, in addition to itself. Therefore, every DMU has n cross
efficiencies calculated from n sets of multipliers selected by n DMUs. The averages
of the n cross efficiencies from every DMU are then used for ranking. Since there are
multiple solutions for using either the CCR or BCC models to measure the efficiency
of every DMU, and improperly selected multipliers can lead to misleading results,
one approach is to select the multipliers that produce the maximum weighted
average efficiency of all DMUs. This procedure uses a conventional DEA model
to calculate the efficiency of a DMU to obtain the efficiency E0. Then the multipliers
that maximize the weighted average efficiency of the n DMUs, while maintaining the
efficiency of this DMU at E0, are sought via the following model:
n s
j=1 r = 1 ur Y rj
max: n m
j=1 i = 1 vi X ij
m
s
s:t: r = 1 ur Y r0 = E0 vi X i0
i=1 ð2:21Þ
s
r = 1 ur Y rj
m ≤ 1, j = 1, . . . , n
i = 1 vi X ij
ur , vi ≥ ε, r = 1, . . . , s, i = 1, . . . , m:
Note that the objective function is the average of the efficiencies

E j = sr = 1 ur Y rj = m
i = 1 vi X ij , j = 1, . . ., n, weighted by the proportion of their
aggregate input in the total aggregate input, m i = 1 vi X ij =
n
j=1
m
i = 1 vi X ij . The opti-

mal solution (u , v ) is then used to calculate the efficiency of every DMU d, Ed0.
This procedure is repeated for every DMU f to obtain the cross efficiency for every
DMU d, Edf. The average efficiency for DMU d, E d = nf = 1 E df =n, is then used for
ranking.
Most of the ranking methods are based on the idea of restricting the range of the
multipliers that are used to calculate the efficiencies. In contrast, Andersen and
Petersen (1993) proposed eliminating the focal DMU to construct the frontier from
the remaining n-1 DMUs in order to calculate an efficiency index for ranking.
However, this method is only for ranking efficient DMUs. Since the DMUs being
eliminated are efficient ones, they will fall outside of the region encompassed by the
new frontier, and their efficiency scores calculated based on this frontier will be
greater than one. This is why this efficiency index is said to measure super efficiency.
The following is an output model for calculating the super efficiency of the focal
DMU under variable returns to scale:
2.4 Ranking 39
Fig. 2.7 Calculation of Y

super efficiency
10
E
8 S’
D
6
B
4
F C
2 A
S G
O 2 4 6 8 10 X
m
1 i = 1 vi X i0 þ v0
= min: s
E0 r = 1 ur Y r0
m
i = 1 vi X ij þ v0
s:t: s ≥ 1, j = 1, . . . , n, j≠0 ð2:22Þ
r = 1 ur Y rj
ur , vi ≥ ε, r = 1, . . . , s, i = 1, . . . , m
Figure 2.7 is a graphical interpretation of super efficiency using the example in

Fig. 2.4, where five DMUs, A, B, C, D, and E, use different amounts of input X to
produce different amounts of output Y. The frontier constructed from these DMUs is
the line segments SABDES’, and DMUs A, B, D, and E are efficient. In order to rank
these four efficient DMUs, they are each eliminated in turn to construct new frontiers
to calculate their super efficiencies. For example, the super efficiency of DMU B is
measured from the frontier SADES’ constructed from the other four DMUs, A, C, D,
and E, as BG/FG. The super efficiencies of DMUs D and E can be calculated
similarly. To calculate the super efficiency of DMU A, however, one will obtain
an unbounded value, because DMU A does not have a line segment with which to
calculate efficiency. This situation will not occur under constant returns to scale.
While several methods have been proposed to solve this problem (Li et al., 2007;
Cook et al., 2009), super efficiency does not seem to be a suitable method for ranking
(Banker & Chang, 2006).
The major difference between the CCR and BCC models is the effect of scale, and
many articles address this issue. Seiford and Zhu (1999) reviewed three basic
methods for determining returns to scale and the effects of multiple solutions.
Banker et al. (2004) discussed returns to scale for several DEA models. Some
other works related to this topic are Jahanshahloo and Soleimani-Damaneh (2004),
Zarepisheh and Soleimani-Damaneh (2009), Fukuyama (2000), and Førsund and
Hjalmarsson (2004).
The issue of weight restrictions is also widely discussed in the DEA literature, and
there are different ways of classifying the related methods. The following studies
have reviewed the literature on this topic: Roll and Golany (1993), Allen et al.
(1997), Angulo-Meza and Estellita Lins (2002), Joro and Viitala (2004), and Sarrico
and Dyson (2004). The works of Podinovski and Thanassoulis (2007), Khalili et al.
(2010), Podinovski and Bouzdine-Chameeva (2013), and Førsund (2013) also
discuss this issue.
Soltanifar and Lotfi (2011) discussed the strengths and weaknesses of several
ranking methods and proposed the voting analytic hierarchy process method. Wu
et al. (2012) compared different cross efficiency methods for ranking. Other com-
prehensive reviews of the methods used for ranking include Adler et al. (2002),
Singh and Chand (2007), Jablonsky (2012), and Hosseinzadeh et al. (2013).
References
Adler, N., Friedman, L., & Sinuany-Stern, Z. (2002). Review of ranking methods in the data
envelopment analysis context. European Journal of Operational Research, 140, 249–265.
Allen, R., Athanassopoulos, A., Dyson, R. G., & Thanassoulis, E. (1997). Weights restrictions and
value judgments in data envelopment analysis: Evolution development and future directions.
Annals of Operations Research, 73, 13–34.
Andersen, P., & Petersen, N. C. (1993). A procedure for ranking efficient units in data envelopment
analysis. Management Science, 39, 1261–1264.
Angulo-Meza, L., & Estellita Lins, M. P. (2002). Review of methods for increasing discrimination
in data envelopment analysis. Annals of Operations Research, 116, 225–242.
Banker, R. D., & Chang, H. H. (2006). The super-efficiency procedure for outlier identification, not
for ranking efficient units. European Journal of Operational Research, 175, 1311–1320.
Banker, R. D., & Thrall, R. M. (1992). Estimation of returns to scale using data envelopment
analysis. European Journal of Operational Research, 62, 74–84.
Banker, R. D., Charnes, A., Cooper, W. W., Swarts, J., & Thomas, D. (1989). An introduction to
data envelopment analysis with some of its models and their uses. Research in Governmental
and Nonprofit Accounting, 5, 125–163.
Banker, R. D., Chang, H. H., & Cooper, W. W. (1996). Equivalence and implementation of
alternative methods for determining returns to scale in data envelopment analysis. European
Journal of Operational Research, 89, 473–481.
References 41
Banker, R. D., Cooper, W. W., Seiford, L. M., Thrall, R. M., & Zhu, J. (2004). Returns to scale in
different DEA models. European Journal of Operational Research, 154, 345–362.
Bitran, G. R., & Chang, L. (1984). Productivity measurement at the firm level. Interfaces, 14(3),
29–40.
Charnes, A., & Cooper, W. W. (1961). Management models and industrial applications of linear
programming. Wiley.
Charnes, A., & Cooper, W. W. (1962). Programming with linear fractionals. Naval Research
Logistics Quarterly, 9, 181–186.
Charnes, A., Cooper, W. W., & Rhodes, E. (1978). Measuring the efficiency of decision making
units. European Journal of Operational Research, 2, 429–444.
Charnes, A., Cooper, W. W., & Rhodes, E. (1979). Short communication: Measuring the efficiency
of decision making units. European Journal of Operational Research, 3, 339.
Charnes, A., & Cooper, W. W. (1984). The non-Archimedean CCR ratio for efficiency analysis: A
rejoinder to Boyd and Färe. European Journal of Operational Research, 15, 333–334.
Charnes, A., Clark, C. T., Cooper, W. W., & Golany, B. (1985). A development study of data
envelopment analysis in measuring the efficiency of maintenance units in the U.S. Air Force.
Charnes, A., Cooper, W. W., & Thrall, R. M. (1986). Classifying and characterizing efficiencies and
inefficiencies in data envelopment analysis. Operations Research Letters, 5, 105–110.
Charnes, A., Cooper, W. W., Wei, Q. L., & Huang, Z. M. (1989). Cone ratio data envelopment
analysis and multi-objective programming. International Journal of Systems Science, 20,
1099–1118.
Charnes, A., Cooper, W. W., & Thrall, R. M. (1991). A structure for classifying and characterizing
efficiencies and inefficiencies in data envelopment analysis. Journal of Productivity Analysis, 2,
197–237.
Cook, W. D., Liang, L., Zha, Y., & Zhu, J. (2009). A modified super-efficiency DEA model for
infeasibility. Journal of the Operational Research Society, 60, 276–281.
Doyle, J., & Green, R. (1994). Efficiency and cross-efficiency in DEA: Derivations, meanings, and
uses. Journal of the Operational Research Society, 45, 567–578.
Dyson, R. G., & Thanassoulis, E. (1988). Reducing weight flexibility in data envelopment analysis.
Journal of the Operational Research Society, 39, 563–576.
Färe, R., & Hunsaker, W. (1986). Notions of efficiency and their reference sets. Management
Science, 32, 237–243.
Førsund, F. R. (2013). Weight restrictions in DEA: Misplaced emphasis? Journal of Productivity
Analysis, 40, 271–283.
Førsund, F. R., & Hjalmarsson, L. (2004). Calculating scale elasticity in DEA models. Journal of
the Operational Research Society, 55, 1023–1038.
Fukuyama, H. (2000). Returns to scale and scale elasticity in data envelopment analysis. European
Hosseinzadeh, L. F., Jahanshahloo, G. R., Khodabakhshi, M., Rostamy-Malkhlifeh, M.,
Moghaddas, Z., & Vaez-Ghasemi, M. (2013). A review of ranking models in data envelopment
analysis. Journal of Applied Mathematics, 2013, 1–20.
Jablonsky, J. (2012). Multicriteria approaches for ranking of efficient units in DEA models. Central
European Journal of Operations Research, 20, 435–449.
Jahanshahloo, G. R., & Soleimani-Damaneh, M. (2004). Estimating returns to scale in data
envelopment analysis: A new procedure. Applied Mathematics and Computation, 150, 89–98.
Joro, T., & Viitala, E. J. (2004). Weight-restricted DEA in action: From expert opinions to
mathematical models. Journal of the Operational Research Society, 55, 814–821.
Kao, C., & Hung, H. T. (2005). Data envelopment analysis with common weights: The compromise
solution approach. Journal of the Operational Research Society, 56, 1196–1203.
Khalili, M., Camanho, A. S., Portela, M. C. A. S., & Alirezaee, M. R. (2010). The measurement of
relative efficiency using data envelopment analysis with assurance regions that link inputs and
outputs. European Journal of Operational Research, 203, 761–770.
Koopmans, T. C. (1951). An analysis of production as an efficient combination of activities. In T. C.

Koopmans (Ed.), Activity analysis of production and allocation. Cowles Commission for
Research in Economics. Monograph No. 13 (pp. 33–97). Wiley.
Lewin, A., & Morey, R. (1981). Measuring the relative efficiency and output potential of public
sector organizations: An application of data envelopment analysis. International Journal of
Policy Analysis and Information Systems, 5, 267–285.
Li, S., Jahanshahloo, G. R., & Khodabakhshi, M. (2007). A super-efficiency model for ranking
efficient units in data envelopment analysis. Applied Mathematics and Computation, 184,
638–648.
Podinovski, V. V. (1999). Side effects of absolute weight bounds in DEA models. European
Podinovski, V. V., & Bouzdine-Chameeva, T. (2013). Weight restrictions and free production in
data envelopment analysis. Operations Research, 61, 426–437.
Podinovski, V. V., & Thanassoulis, E. (2007). Improving discrimination in data envelopment
analysis: Some practical suggestions. Journal of Productivity Analysis, 28, 117–126.
Roll, Y., & Golany, B. (1993). Alternate methods of treating factor weights in DEA. Omega, 21,
99–109.
Roll, Y., Cook, W. D., & Golany, B. (1991). Controlling factor weights in data envelopment
analysis. IIE Transactions, 23, 2–9.
Sarrico, C. S., & Dyson, R. G. (2004). Restricting virtual weights in data envelopment analysis.
Seiford, L. M., & Zhu, J. (1999). An investigation of returns to scale in data envelopment analysis.
Omega, 27, 1–11.
Singh, O., & Chand, S. (2007). Ranking of decision making units: A review and development of
new model using data envelopment analysis approach. Opsearch, 44, 185–201.
Soltanifar, M., & Lotfi, F. H. (2011). The voting analytic hierarchy process method for discrimi-
nating among efficient decision making units in data envelopment analysis. Computers and
Industrial Engineering, 60, 585–592.
Thompson, R. G., Singleton, F. D., Jr., Thrall, R. M., & Smith, B. A. (1986). Comparative site
evaluations for locating a high-energy physics lab in Texas. Interfaces, 16(6), 35–49.
Thompson, R. G., Langemeier, L. N., Lee, C. T., Lee, E., & Thrall, R. M. (1990). The role of
multiplier bounds in efficiency analysis with applications to Kansas farming. Journal of
Econometrics, 46, 93–108.
Tracy, D. L., & Chen, B. (2005). A generalized model for weight restrictions in data envelopment
analysis. Journal of the Operational Research Society, 56, 390–396.
Wong, Y. B., & Beasley, J. E. (1990). Restricting weight flexibility in data envelopment analysis.
Wu, J., Sun, J., & Liang, L. (2012). Cross efficiency evaluation method based on weight-balanced
data envelopment analysis. Computers and Industrial Engineering, 63, 513–519.
Zarepisheh, M., & Soleimani-Damaneh, M. (2009). A dual simplex-based method for determina-
tion of the right and left returns to scale in DEA. European Journal of Operational Research,
194, 585–591.
Chapter 3
Distance Function Efficiency Measures
While management scientists were developing methods to measure efficiency based

on the definition of productivity, economists were tackling the same problem using
the production function. In the former, efficiency is measured as the ratio of
aggregate output to aggregate input, which, from the input point of view, is equal
to the smallest amount of input required to produce a given amount of output divided
by the actual amount of input consumed. This idea is similar to defining an input
distance function to measure the relative distance between the minimum input and
the actual input as the input efficiency. From the output point of view, the output–
input ratio measure of efficiency is equal to the actual amount of output produced
divided by the largest amount of output that can be produced from a given amount of
input. An output distance function can thus be defined to measure the relative
distance between the actual and maximum outputs as the output efficiency.
In Chap. 1 it was mentioned that any dominating point of a DMU on the
production function can be selected as the benchmark to measure the efficiency of
this DMU. The output–input ratio approach, however, is able to use only two points,
one from the input side and the other from the output side, as the benchmark to
measure efficiency. By defining a directional distance function, any dominating
point can be selected as the benchmark to measure efficiency. The former is a special
case of the latter, and since its direction for measuring efficiency is a ray passing
through the origin (in either the input or output space), it is a radial measure. The
latter, which does not necessarily pass through the origin, is a non-radial measure.
This chapter starts with a discussion of the production possibility set, followed by the
input and output distance functions, and finishes with the directional distance
function.

https://doi.org/10.1007/978-3-031-27593-7_3
44 3 Distance Function Efficiency Measures
3.1 Production Possibility Set
Let H(X, Y) = H(X1, . . ., Xm, Y1, . . ., Ys) = 0 be the implicit form of the production
function, where m inputs are used to produce s outputs. Any unit on the production
frontier with H(X, Y) = 0 is efficient, and that consuming more inputs than required
to produce the same amount of outputs, or consuming the same amount of inputs to
produce less amount of outputs, with H(X, Y) < 0, is inefficient. Theoretically, all
inefficient units exist, and they constitute the production possibility set T, which is
defined as:
T = fðX, Y Þ j Y ≥ 0 can be produced by X ≥ 0g ð3:1Þ
Curve OR in Fig. 3.1 represents a production function H(X, Y) = 0. The shaded

region encompassed by this curve is the production possibility set associated with
this production function.
When the s outputs are fixed at a set of positive values, represented by the vector
Y0, one has an isoquant of the form q(X) = Y0. All points on the isoquant are
efficient, and all points dominated by the isoquant are inefficient. Domination means
consuming more units of at least one input to produce the same amount of output Y0.
Theoretically, all inefficient points exist, and they form the input possibility set
Q(Y0) for the given output Y0:
QðY 0 Þ = fX j X ≥ 0 is able to produce Y 0 g ð3:2Þ
Fig. 3.1 Production Y

function and production
possibility set
R
C
H(X,Y)=0
Y0 A B
Q(Y0)
T
P(X0)
D
O X0 X
3.1 Production Possibility Set 45
Fig. 3.2 Input possibility X2

set and input distance
function
S
Q(Y0)
q(X)=Y0 A
B=θ*A
O X1
In other words, all points on the isoquant q(X) = Y0 and those dominated by the
isoquant constitute the input possibility set Q(Y0). Ray AB in Fig. 3.1 is the input
possibility set for output level Y0.
If the points in the input possibility set are able to operate efficiently, then they
should be able to produce more units of at least one output than Y0 to result in
q(X) ≥ Y0. Note that q(X) has a domain of X in Rm and a range of Y in Rs . The
relationship of Y ≥ Y0 means that every corresponding component of the former is
greater than or equal to the latter, with at least one component being strictly greater
than the other. Curve SS′ in Fig. 3.2 represents the isoquant q(X) = Y0 in the input
space. The shaded region to the top right of the isoquant, with q(X) ≥ Y0, is the input
possibility set Q(Y0).
In contrast, if the m inputs are fixed at a set of values, denoted as X0, then one has
a product transformation curve of the form p(Y) = X0. All units on this curve are
efficient, and those dominated by this curve are inefficient. Again, it is assumed that
all inefficient units exist theoretically, and they constitute the output possibility set
p(X0) for input level X0. Mathematically, the output possibility set is defined as:
pðX 0 Þ = fY j Y ≥ 0 can be produced by X0 g ð3:3Þ
The points in the output possibility set are dominated by the product transforma-
tion curve p(Y) = X0. The line segment CD in the X-Y space of Fig. 3.1 is the output
possibility set for input level X0. Figure 3.3 shows a product transformation curve
p(Y) = X0 in the output space. The shaded region is the output possibility set for
input level X0. Under perfect conditions, the points in the output possibility set
should be produced with less amount of X0. One thus has p(Y) ≤ X0 for Y 2 P(X0).
Fig. 3.3 Output possibility Y2

set and output distance
function
B=ϕA
•A
p(Y)=X0
T
O Y1
The above discussion is based on a theoretical production function H(X, Y) = 0.

In practice, one needs a sample of points to construct an empirical production
function. Given a set of n DMUs, with observations of (Xj, Yj) = (X1j, . . ., Xmj,
Y1j, . . ., Ysj), j = 1, . . ., n Banker et al. (1984) postulated convexity and dispos-
ability to construct the production possibility set T.
n n
Convexity: If (Xj, Yj) 2 T, j = 1, . . ., n, then j = 1 λj X j , j = 1 λj Y j 2 T, for λj ≥ 0,
n
j = 1, . . ., n, and j = 1 λj
= 1.
Disposability: If (X0, Y0) 2 T, then (X, Y0) 2 T for X ≥ X0, and (X0, Y) 2 T for Y ≤ Y0.
Convexity states that the convex combinations of the n DMUs also belong to the
production possibility set. Disposability indicates that inefficient production is
allowed and is expressed in the forms of consuming more amount of at least one
input to produce the same amount of outputs, consuming the same amount of inputs
to produce smaller amount of at least one output, or both. Since all n DMUs belong
n n n
to the production possibility set T, j = 1 λj X j , j = 1 λj Y j , with j = 1 λj = 1, λj ≥ 0,
j = 1, . . ., n, by convexity, also belong to T. Furthermore, (X, Y) belongs to T for
X ≥ nj = 1 λj X j and Y ≤ nj = 1 λj Y j , by disposability. Combining these two properties,
the production possibility set constructed from the n DMUs is:
3.2 Input Distance Function 47
T
n n n
= ðX, Y Þj λj X j ≤ X, λj Y j ≥ Y, λj = 1, λj ≥ 0, j = 1, . . . , n
j= =1 j= =1 j= =1
ð3:4Þ
It should be noted that different postulates define different production possibility

sets. For example, if the returns to scale of production are assumed to be constant,
then the postulate of ray unboundedness should be further considered (Banker et al.,
1984):
Ray unboundedness: If (X0, Y0) 2 T, then (cX0, cY0) 2 T, for c > 0.
With this postulate the conditions for a point to belong to the production
possibility set are modified to: c nj = 1 λj Xj ≤ X, c nj = 1 λj Y j ≥ Y, n
j = 1 λj = 1,
λj ≥ 0, j = 1, . . ., n, and c > 0. By replacing cλj with μj, the conditions become
n n n
j = 1 μj X j ≤ X, j = 1 μj Y j ≥ Y, j = 1 μj = c, μj ≥ 0, j = 1, . . ., n, and c > 0. Since
c can be any number such that the constraint nj = 1 μj = c is always satisfied, this
constraint can be deleted to result in the following expression for the production
possibility set TCRS under constant returns to scale (CRS):
n n
T CRS = ðX, Y Þ j λj Xj ≤ X, λj Y j ≥ Y, λj ≥ 0, j = 1, . . . , n ð3:5Þ
j= =1 j= =1
The production possibility sets constructed under variable and constant returns to
scale are quite similar. The only difference is that the former needs the convexity
constraint of nj = 1 λj = 1, while the latter does not.
3.2 Input Distance Function
Let ||X|| be the norm of vector X. Shephard (1953, 1970) defined a distance function
Φ(Y, X) on the input possibility set Q(Y) as:
X 1
ΦðY, X Þ = = ð3:6Þ
θðY, X Þ X θðY, X Þ
where
θðY, X Þ = min:fθ j θX 2 QðY Þ, θ ≥ 0g ð3:7Þ
Given a point (X0, Y0), its input distance measure Φ(Y0, X0), according to
Eqs. (3.6) and (3.7), is the inverse of parameter θ, where θ is the smallest contraction
of (X0, Y0) such that the point (θX0, Y0) is still in the input possibility set Q(Y0).
Consider point A in Fig. 3.2. For θ ≥ 0, θA is the ray OA, where the origin
corresponds to θ = 0, and point A corresponds to θ = 1. Any point on the line
segment OA has a positive θ less than or equal to one, and point B on the isoquant has
the smallest value θ such that the point θA is still in the input possibility set Q(Y0).
Suppose the production possibility set T is constructed from a set of n DMUs
under constant returns to scale, as defined by Expression (3.5), Model (3.7) for
measuring the distance of a DMU indexed by 0 is a linear program, which can be
formulated as:
min: θ
n
s:t: λj X ij ≤ θX i0 , i = 1, . . . , m
j=1
n
ð3:8Þ
λj Y rj ≥ Y r0 , r = 1, . . . , s
j=1
λj ≥ 0, j = 1, . . . , n
θ ≥ 0:
This model is the same as the envelopment form of the CCR ratio model (2.4) for
measuring efficiency under constant returns to scale introduced in Chap. 2, except
that the non-Archimedean number ε is ignored and parameter θ is required to be
positive. In other words, the Shephard input distance function value for the DMU,
Φ(Y0, X0), is just the inverse of the input efficiency of this DMU. Referring to DMU
A in Fig. 3.2, one has θ = OB/OA, which is the same as that of the output–input ratio
measure of efficiency, and point B is the benchmark.
Model (3.8) requires the contraction parameter θto be positive. Since λj ≥ 0, j = 1,
. . ., n, which result in nj = 1 λj X ij ≥ 0, the constraint nj = 1 λj X ij ≤ θX i0 ensures that θ
is positive. The non-negativity constraint θ ≥ 0 can thus be deleted. Furthermore,
Model (3.8) ignores the effect of the non-Archimedean number. In this case, the
weakly efficient DMUs will be misjudged as efficient. To be Pareto efficient a DMU
is required to have θ = 1 and zero for all slack variables at optimality. These two
conditions modify Model (3.8) to the following form, which is exactly the same as
Model (2.4):
m s
E0 = min: θ-ε si- þ sþ
r
i=1 r=1
n
j=1
n
λj Y rj - sþ
r = Y r0 , r = 1, . . . , s
j=1
λj , si- , sþ
r ≥ 0, j = 1, . . . , n, i = 1, . . . , m, r = 1, . . . , s
θ unrestricted in sign,
ð3:9Þ
where E0 is the efficiency of the DMU being evaluated, and it is equal to one only if
θ = 1 and si- = sþ
r = 0, i = 1, . . ., m, r = 1, . . ., s.
The benchmark for measuring the efficiency of this DMU, or the target for this
n n
DMU to become efficient, is j = 1 λj X ij , j = 1 λj Y rj , where
n
λj X ij = θ X i0 - si- , i = 1, . . . , m
j=1
n
ð3:10Þ
λj Y rj = Y r0 þ sþ
r , r = 1, . . . , s
j=1
are obtained from the constraints of Model (3.9). To become efficient, this DMU
must reduce each input Xi0 to θ X i0 - si- , and increase each output Yr0 by sþ
r .
However, it should be noted that any point on the production frontier can be chosen
as a benchmark for measuring efficiency, and a target for making improvement.
The distance function can also be defined on the input possibility set under
variable returns to scale, as expressed by Expression (3.4). The inverse of the
distance function, taking the non-Archimedean number ε into consideration and
ignoring the positivity requirement of θ, can be expressed as:
m s
E0 = min: θ - ε si- þ sþ
r
i=1 r=1
n
j=1
n
λj Y rj - sþ
r = Y r0 , r = 1, . . . , s
j=1
n
λj = 1
j=1
λj , si- , sþ
r ≥ 0, j = 1, . . . , n, i = 1, . . . , m, r = 1, . . . , s
ð3:11Þ
This model is exactly the same as Model (2.10), the envelopment form of the
BCC input model for measuring efficiency. The target for making improvement is
the same as that shown in Eq. (3.10). The only difference is that the linear combi-
nation variables λj are required to have a sum of one.
The DMUs corresponding to the linear combination variables λj with nonzero
values are the reference points, and their convex combinations constitute the frontier
facet for the DMU being assessed to measure efficiency. The sum of the nonzero λj
can be used to judge the status of returns to scale of the focal DMU. Banker and
Thrall (1992) showed that increasing, constant, and decreasing returns to scale
prevail when nj = 1 λj < 1, nj = 1 λj = 1, and nj = 1 λj > 1, respectively, provided
the DMU being evaluated is technically efficient. Banker et al. (1996) later elimi-
nated this requirement in a formulation to identify returns to scale. The efficiency
calculated from Model (3.11), under variable returns to scale, is technical efficiency,
that from Model (3.9), under constant returns to scale, is the overall efficiency, and
the ratio of the latter to the former is the scale efficiency.
In order to compare with the output–input ratio measures discussed in Chap. 2,
the data of the example in Table 2.1, which is duplicated in Table 3.1 for easy
reference, is used for illustration. Figure 2.4 is also redrawn here as Fig. 3.4. By
applying Model (3.9), the efficiencies of the five DMUs and the associated optimal
nonzero λj variables are shown in columns five and six of Table 3.1. Since all slack
variables are zero for every DMU, the distance measure θ is just the efficiency.
Every DMU has only one nonzero linear combination variable λj, indicating that the
corresponding DMU j is the reference point. Specifically, DMU A has λB = 0:5, its
reference point is thus DMU B. Its benchmark, according to Eq. (3.10), is point S,
with coordinates λB X B , Y A = ð2, 2Þ. We already know that the region with a scale
smaller than that of DMU B is of increasing returns to scale, DMU A, with
n
j = 1 λj = λB = 0:5 < 1, is thus in the region of increasing returns to scale. For
DMUs B and D, one has λB = λD = 1, hence both the reference point and the
3.2
Input Distance Function
Table 3.1 Input distance measures of an example of one input and one output
Scale
Constant returns to scale Variable returns to scale eff.
DMU X Y E0 θ λj Benchmark θ λj Benchmark
A 3 2 2/3 2/3 λB = 0:5 S = (2, 2) 1 λA = 1 A = (3, 2) 2/3
B 4 4 1 1 λB = 1 B = (4, 4) 1 λB = 1 B = (4, 4) 1
C 5 3 0.6 0.6 λB = 0:75 U = (3, 3) 0.7 λA = λB = 0:5 W = (3.5, 3) 6/7
D 6 6 1 1 λD = 1 D = (6, 6) 1 λD = 1 D = (6, 6) 1
E 10 8 0.8 0.8 λD = 4=3 R = (8, 8) 1 λE = 1 E = (10, 8) 0.8
51
Fig. 3.4 Input and output Y

distance measures in the
X-Y space Z
10
8 R
E
6
D
V : YV = 53 YC
B
4
(XU =0.6XC)U C
W: XW =0.7XC
2 S A
O 2 4 6 8 10 X
benchmark are themselves, and they are located in the region of constant returns to
scale, as nj = 1 λj = λB = λD = 1. DMU C has λB = 0:75, with DMU B as its refer-
ence point, and U = λB X B , Y C = ð3, 3Þ as its benchmark. Since U is in the region of
increasing returns to scale, DMU C is in this region as well. Finally, DMU E has
λD = 4=3, indicating that its reference point is DMU D, and it is located in the region
of decreasing returns to scale, because nj = 1 λj = λD = 4=3 > 1: Its benchmark is
R = λD X D , Y E = ð8, 8Þ. Column seven of Table 3.1 shows the benchmarks of the
five DMUs. Notably, when the computer software LINGO was used to solve this
example, the optimal linear combination variable obtained was λB = 1:5 for DMU
D and λB = 2 for DMU E, although θ was still the same for these two cases, a
situation with multiple solutions.
Under variable returns to scale Model (3.11) is applied to measure efficiencies.
For the data of the five DMUs contained in Table 3.1, the efficiencies obtained are
1, 1, 0.7, 1, and 1, as shown in the fourth-to-last column, where only DMU C is
inefficient. All slack variables are zero. For the efficient DMUs the reference points
and benchmarks are themselves, while for the inefficient DMU C the reference
points are DMUs A and B, with λA = λB = 0:5, and the benchmark is
λA A þ λB B = ð3:5, 3Þ = W. The ratio of the efficiency under constant returns to
scale to that under variable returns to scale is scale efficiency, as shown in the last
column of Table 3.1. All these results are the same as those measured from the
output–input ratio models.
Consider another example in Table 2.2, with the contents copied in Table 3.2 to
maintain continuity. In this example, five DMUs use different amounts of inputs X1
and X2 to produce one unit of output Y. Figure 3.5 is a replication of Fig. 2.2. By
Table 3.2 Input distance measures of an example of two inputs and one output
DMU X1 X2 Y E0 θ* λj Slack Benchmark
A 1 4 1 1 1 λA = 1 0 A
B 2 2 1 1 1 λB = 1 0 B
C 4 1 1 1 1 λC = 1 0 C
D 4 3 1 0.6 0.6 λB = 0:8, λC = 0:2 0 U
E 6 1 1 1-2ε 1 λC = 1 s1- = 2 C
Fig. 3.5 Input and X2

directional distance
measures in the input space
4 A
F
D
2 B G=0.6D
C E
O 2 4 6 X1
applying Model (3.9) of constant returns to scale, the optimal solutions of nonzero
variables are shown in the fourth- to the second-to-last columns of Table 3.2.
DMUs A, B, and C have θ = 1 and zero for all slack variables, indicating that
they are efficient. The reference points and benchmarks are themselves, as indicated
by the only nonzero linear combination variables of λA = 1, λB = 1, and λC = 1,
respectively. For DMU D the reference points are DMUs B and C, because the
nonzero linear combination variables are λB = 0:8 and λC = 0:2. The benchmark is
the linear combination of λB B þ λC C = ð2:4, 1:8Þ = G. Since all slack variables are
zero, the efficiency of DMU D is the distance measure θ, which is 0.6. The optimal
solution for DMU E is θ = 1, s1- = 2, and λC = 1, which indicates that it is weakly
efficient, with an efficiency score of 1-2ε, and both the reference point and
benchmark are DMU C.
3.3 Output Distance Function
In addition to defining a distance function on the input possibility set, Shephard

(1953, 1970) also defined a distance function Ω(X, Y) on the output possibility set
P(X) as:
Y 1
ΩðX, Y Þ = = ð3:12Þ
φðX, Y Þ Y φðX, Y Þ
where
φðX, Y Þ = max:fφ jφ Y 2 PðX Þ, φ ≥ 0g ð3:13Þ
For a set of n DMUs, with observations (Xj, Yj), j = 1, . . ., n, the corresponding

production possibility set under constant returns to scale, according to Expression (3.5),
is T CRS = ðX, Y Þj nj = 1 λj X j ≤ X, n
j = 1 λj Y j ≥ Y, λj ≥ 0, j = 1, . . . , n , and
the output possibility set for input vector X0 is P(X0) = {Y | Y ≥ 0 can be produced by
X0}. Referring to Fig. 3.3, the output possibility set is the shaded area. For DMU A, φΑ,
for φ ≥ 0 is the ray OA. The origin has φ = 0, DMU A has φ = 1, and point B on the
product transformation curve TT′ has φ ≥ 1.
Model (3.13) searches for the largest value of φ such that φY is still in the output
possibility set P(X0). To measure the output distance of a DMU, φ(X0, Y0), defined in
Model (3.13), becomes:
max: φ
n
s:t: λj X ij ≤ X i0 , i = 1, . . . , m
j=1
n
ð3:14Þ
λj Y rj ≥ φY r0 , r = 1, . . . , s
j=1
λj ≥ 0, j = 1, . . . , n
φ ≥ 0:
This model is the same as the envelopment form of the CCR output ratio model
(2.7), except that the non-Archimedean number ε has been ignored, and the distance
measure φ is required to be positive. Since λj ≥ 0, j = 1, . . ., n, one has
n n
j = 1 λj Y rj ≥ 0, which requires φ ≥ 0 to satisfy the constraint of j = 1 λj Y rj ≥ φY r0
in Model (3.14) and makes the constraint of φ ≥ 0 unnecessary. As a matter of fact, φ
is an expansion factor for the DMU being evaluated to project onto the product
transformation curve, and it will always be greater than or equal to one. Regarding
the non-Archimedean number ε, since this DMU will be Pareto efficiency only if
3.3 Output Distance Function 55
φ = 1 and all slack variables equal zero, the objective function can be modified to
m - s þ
“ max:φ þ ε i = 1 si þ r = 1 sr ” to obtain the following model:
m s
max: φþε si- þ sþ
r
i=1 r=1
n
s:t: λj X ij þ si- = X i0 , i = 1, . . . , m
j=1
n ð3:15Þ
λj Y rj - sþ
r = φY r0 , r = 1, . . . , s
j=1
λj , si- , sþ
r ≥ 0, j = 1, . . . , n, i = 1, . . . , m, r = 1, . . . , s
φ unrestricted in sign,
which is exactly the same as Model (2.7). The objective value is the inverse of the
efficiency of the DMU being evaluated, and the output distance function calculated
at (X0, Y0), Ω(X0, Y0) is just the efficiency of this DMU.
The benchmark for this DMU to measure efficiency is the point (X0, φY0). When
the slack variables are taken into consideration, the benchmark becomes:
n
λj X ij = X i0 - si- , i = 1, . . . , m
j=1
n
ð3:16Þ
λj Y rj
= φ Y r0 þ sþ
r , r = 1, . . . , s
j=1
The benchmark also serves as a target for this DMU to become efficient, by
reducing the input Xi0 by si- and bringing the output Yr0 to the level of φ Y r0 þ sþ r .
When the assumption of constant returns to scale is relaxed to allow for variable
returns to scale, the production possibility set becomes
T = ðX, Y Þj nj = 1 λj X j ≤ X, nj = 1 λj Y j ≥ Y, nj = 1 λj = 1, λj ≥ 0, j = 1, . . . , n . The
output distance measure of Model (3.13), taking the non-Archimedean number ε
into consideration and ignoring the positivity constraint for φ, can be formulated as:
m s
r
i=1 r=1
n
s:t: λj X ij þ si- = X i0 , i = 1, . . . , m
j=1
n
λj Y rj - sþ
r = φY r0 , r = 1, . . . , s ð3:17Þ
j=1
n
λj = 1
j=1
λj , si- , sþ
r ≥ 0, j = 1, . . . , n, i = 1, . . . , m, r = 1, . . . , s
φ unrestricted in sign,
which is exactly the same as Model (2.14), the envelopment form of the BCC output
ratio model for measuring the efficiency of this DMU. The target point can be
obtained via an equation the same as Eq. (3.16).
Consider again the one-input one-output example in Table 3.1, with the first four
columns copied in Table 3.3. By applying Model (3.15), the optimal output distance
measure φ and the nonzero linear combination variables λj for the five DMUs are
obtained as shown in columns five and six of Table 3.3. Since all slack variables
have a value of zero, the distance measure φ is just the inverse of the efficiency of
the DMU. The nonzero λj indicates that the reference points for the five DMUs are
B, B, (B and D), D, and D. The sums of the nonzero λj for DMUs A, B, C, D, and
E are 0.75 (<1), 1, 1, 1, and 5/3 (>1), respectively, indicating that DMU A is in the
region of increasing returns to scale, DMUs B, C, and D are in the region of constant
returns to scale, and DMU E is in the region of decreasing returns to scale. The
benchmarks (X0, φY0) are shown in column seven of Table 3.3. Similar to the case
of input distance measures, there are alternative solutions. When LINGO was used to
solve this example, the optimal linear combination variables obtained were λB =
1:25 for DMU C, λB = 1:5 for DMU D, and λB = 2:5 for DMU E, although their φ
values were the same. Figure 3.4 shows the relationships between the benchmarks
and λj values.
Model (3.17) is used to measure the efficiency under variable returns to scale.
With the data shown in Table 3.3, the efficiencies are calculated as shown in the
fourth-to-last column. Since all slack variables are zero, the optimal distance mea-
sure φ of each DMU is equal to the inverse of their efficiency, and there is only
DMU C which is inefficient. The nonzero λj are shown in the third-to-last column,
which also point out the reference points of the corresponding DMU. The bench-
marks of the efficient DMUs, A, B, D, and E, are themselves, while for the inefficient
DMU C it is λΒΒ + λDD = (5, 5) = V. The last column shows the scale efficiency of
every DMU. The geometric relationships are shown in Fig. 3.4.
3.3
Output Distance Function
Table 3.3 Output distance measures of an example with one input and one output
Constant returns to scale Variable returns to scale Scale
*
DMU X Y E0 φ λj Benchmark φ* λj Benchmark eff.
A 3 2 2/3 3/2 λB = 0:75 U = (3, 3) 1 λA = 1 Α = (3, 2) 2/3
B 4 4 1 1 λB = 1 B = (4, 4) 1 λB = 1 B = (4, 4) 1
C 5 3 3/5 5/3 λB = λD = 0:5 V = (5, 5) 5/3 λB = λD = 0:5 V = (5, 5) 1
D 6 6 1 1 λD = 1 D = (6, 6) 1 λD = 1 D = (6, 6) 1
E 10 8 4/5 5/4 λD = 5=3 Z = (10, 10) 1 λE = 1 E = (10, 8) 4/5
57
Table 3.4 Output distance measures of an example with one input and two outputs
DMU X Y1 Y2 E0 φ* Slack λj Benchmark
A 1 2 3 2/3 3/2 0 λB = λD = 0:5 G
B 1 2 5 1 1 0 λB = 1 B
C 1 4 1 1/(1.25 + 0.75ε) 1.25 sþ
2 = 0:75 λE = 1 E
D 1 4 4 1 1 0 λD = 1 D
E 1 5 2 1 1 0 λE = 1 E
Fig. 3.6 Output and Y2

directional distance
measures in the output space
6
B
G=1.5A
4 D
F
A
2 E
O 2 4 6 Y1
For the one-input two-output example in Table 2.3, we copy the content into
Table 3.4 and redraw Fig. 2.3 as Fig. 3.6. The optimal solutions for the five DMUs
under constant returns to scale, as calculated from Model (3.15), are listed in the last
four columns of Table 3.4. There are several things to be noted here. First, except
that DMU C has a nonzero slack of sþ 2 = 0:75, all DMUs have zero slacks, in which
case the efficiency is the reciprocal of φ. The optimal distance measure indicates
how far the DMU can be expanded and still remains in the output possibility set. For
example, DMU A has φ = 1.5, and the farthest position that DMU A can go is point
G, thus φ = OG/OA. Second, the nonzero linear combination variables λj indicate
that the reference points of the five DMUs are (B and D), B, E, D, and E. Third, the
benchmarks for DMUs A, B, C, D, and E, according to Expression (3.16), are G, B,
E, D, and E, respectively. For efficient DMUs the reference points and benchmarks
are themselves.
The above discussions show that the distance function and output–input ratio
approaches produce the same efficiency measures, due to the primal-dual relation-
ship of their corresponding models. However, the distance function approach is
3.4 Directional Distance Function 59
more informative, because it also shows the target points for the inefficient DMUs to
make improvements.
3.4 Directional Distance Function
The input distance function fixes the outputs at the current level to find the smallest
proportion of the inputs that are able to produce the current level of outputs, while
the output distance function fixes the inputs at the current level to find the largest
expansion of the outputs that can be produced from the current level of inputs. As a
matter of fact, neither the inputs nor the outputs need to be fixed in measuring
efficiencies.
Chambers et al. (1996, 1998) defined the directional distance function on the
production possibility set T as:
ηðX, Y, f , gÞ = max:fηjðX - ηf , Y þ ηgÞ 2 T g ð3:18Þ
where f and g are pre-specified directions. For a set of n DMUs with the production
possibility set T defined under variable returns to scale of Expression (3.4), the
directional distance function measured for a DMU along the direction (-f, g) is then:
max: η
n
s:t: λj X ij ≤ X i0 - ηf i , i = 1, . . . , m
j=1
n
λj Y rj ≥ Y r0 þ ηgr , r = 1, . . . , s
j=1 ð3:19Þ
n
λj = 1
j=1
λj ≥ 0, j = 1, . . . , n
η unrestricted in sign:
This function measures the distance from this DMU along the direction (-f, g) to
the point on the production frontier. It shows the amount of outputs that can be
expanded and that of inputs that can be contracted. At optimality, an objective value
of η = 0 indicates that this DMU is efficient, and n
j = 1 λj X ij ,
n
j = 1 λj Y rj shows
the benchmark that this DMU is based on measuring efficiency.
The constraint of nj = 1 λj = 1 in Model (3.19) can be deleted to accommodate
constant returns to scale. The slack variables can also be added to detect weak
efficiency to have the following model:
m s
max: ηþε si- þ sþ
r
i=1 r=1
n
s:t: λj X ij þ si- = X i0 - ηf i , i = 1, . . . , m
j=1
n
λj Y rj - sþ
r = Y r0 þ ηgr , r = 1, . . . , s ð3:20Þ
j=1
n
λj = 1
j=1
λj , si- , sþ
r ≥ 0, j = 1, . . . , n, i = 1, . . . , m, r = 1, . . . , s
Consider DMU C in Fig. 3.7. The distance from this DMU along direction (-f, g)
to the frontier measured by Model (3.20) is η. From this figure it is also clear that if
(-f, g) is set to (-XC, 0), then one obtains η = CW/CL to have an input efficiency of
(1 - η), as measured from the input distance function. Similarly, if (-f, g) is set to
(0, YC), then η = CV/CM is obtained to have an output efficiency of 1/(1 + η), as
measured from the output distance function. In this regard, the directional distance
function is a generalization of Shephard’s distance functions. This feature can
actually be derived rigorously from Model (3.20). Let (-f, g) = (-X0, 0). The
right-hand sides of the constraints in Model (3.20) then become (1 - η)Xi0 and
Yr0. Compared with the input distance model (3.11), one has θ = (1 - η) as the input
efficiency. From the output side, by setting (-f, g) = (0, Y0) to obtain the right-hand
sides of Xi0 and (1 + η)Yr0, the output distance model (3.17) then indicates that the
output efficiency is 1/φ = 1/(1 + η).
m s
min: vi X i0 þ v0 - ur Y r0
i=1 r=1
m s
s:t: vi X ij þ v0 - ur Y rj ≥ 0, j = 1, . . . , n
i=1 r=1
ð3:21Þ
m s
vi f i þ ur gr = 1
i=1 r=1
ur , vi ≥ ε, r = 1, . . . , s, i = 1, . . . , m
The objective function is the difference between the virtual input and the virtual
output, and they are expected to be as close as possible to have a higher efficiency.
An efficient DMU has the same virtual input and output. When (f, g) = (X0, 0),
m
one has i = 1 vi X i0 = 1 from the unity constraint, to obtain an equivalent
3.4 Directional Distance Function 61
Fig. 3.7 Directional Y

distance measures in the X-Y
space
10
8
E
6
D
V
B
4 P η
L C
W
2 A
M Q
O 2 4 6 8 10 X
m s m
objective function of i = 1 vi X i0 þ v0 - r = 1 ur Y r0 = i = 1 vi X i0 = 1 -
s m
r = 1 ur Y r0 - v0 = i = 1 vi X i0 , which, by comparing with the BCC input
model (2.8), is equal to 1 - E0. Similarly, when (f, g) is set to (0, Y0), one
s
has r = 1 ur Y r0 = 1 to obtain the objective function of
m s s m s
ð i = 1 vi X i0 þ v0 - r = 1 ur Y r0 Þ= r = 1 ur Y r0 = ½ð i = 1 vi X i0 þ v0 Þ= r = 1 ur Y r0 - 1,
which, by comparing with the BCC output model (2.12), is equal to (1/E0) - 1. In
other words, the dual formulation also shows that the directional distance measure
is a generalization of the radial distance measure of efficiency.
To further investigate the difference between the measures obtained from the
directional and radial distance functions, first, consider the one-input one-output
example in Table 3.1, with the production frontier shown in Fig. 3.7. Of the five
DMUs, only DMU C is inefficient. The input distance measure, which is equivalent
to the directional distance measure with (-f, g) = (-XC, 0), has a corresponding
η = CW/CL = 0.3, and the output distance measure, which is equivalent to the
directional distance measure with (-f, g) = (0, YC), has a corresponding η = CV/
CM = 2/3. If we set (-f, g) = (-XC, YC), then the benchmark is (3.85, 3.69), or
(50/13, 48/13) to be exact, with η = CP/CQ = 0.23 or 3/13, and λA = 2=13,
λB = 11=13:
For the two-input one-output example in Table 3.2, consider three cases,
( f1, f2, g) = (-X1D, 0, 0), ( f1, f2, g) = (0, -X2D, 0), and ( f1, f2, g) = (-X1D, -X2D, 0).
The benchmarks for DMU D, as shown in Fig. 3.5, are F = (1.5, 3), C = (4, 1), and
G = (2.4, 1.8), with η = 5/8, η = 2/3, and η = 0.4, respectively, where the third case
corresponds to the input distance measure θ, with η = GD/OD = 1 - θ = 0.4.
Similarly, in the one-input two-output example in Table 3.4, the directions of
( f, g1, g2) = (0, Y1A, 0), ( f, g1, g2) = (0, 0, Y2A), and ( f, g1, g2) = (0, Y1A, Y2A), for
DMU A have benchmarks of F = (4.5, 3), B = (2, 5), and G = (3, 4.5), as
shown in Fig. 3.6, with η = AB = 2/3, η = AF = 1.25, and η = AG = 0.5,
respectively, where the third case corresponds to the output distance measure φ,
with η = AG/AO = φ - 1 = 0.5.
To summarize, this chapter shows that the Shephard’s distance function (also
referred to as radial distance function) measures are the same as the output–input
ratio measures discussed in Chap. 2. Furthermore, the directional distance function
measures are generalizations of the Shephard’s distance function measures.
The production possibility sets discussed in this chapter are convex. Another
production possibility set, which is not convex, is the Free Disposable Hull (FDH)
defined as T FDH =fðX,YÞ j nj=1 λj X j ≤X, nj=1 λj Y j ≥Y, nj=1 λj =1,λj 2 f0,1g,j=1, ...,ng.
The difference is that λj is allowed to have only two values, 0 and 1, rather than
positive values. Discussions of this approach for measuring efficiency can be found
in Cherchye et al. (2001), Deprins et al. (1984), Kuosmanen (2001), Leleu (2009),
Soleimani-damaneh et al. (2006), Soleimani-damaneh and Mostafaee (2009), Thrall
(1999), and Tulkens (1993). The idea of the directional distance function can be
traced back to the shortage function of Luenberger (1992, 1995). It is a generaliza-
tion of the gauge function of McFadden (1978), the Farrell proportional distance
function of Briec (1997), as well as the radial distance function of Shephard
discussed in this chapter. In the context of efficiency measurement, the duality result
implies that this orientation of measurement is consistent with the notion of profit
efficiency presented in Nerlove (1965). Both the radial and directional distance
functions have been used to measure the allocative efficiency discussed in
Chap. 1. There are numerous articles addressing this issue, and the following are
some representative ones: Aparicio et al. (2013), Badunenko et al. (2008), Banker
et al. (2007), Chavas and Cox (1999), Fukuyama and Weber (2003, 2005), Sahoo
et al. (2014), Sueyoshi (1992), Zelenyuk (2013), and Zofio et al. (2013). Finally,
there are other distance functions, for example, the Hölder distance function (Briec,
1999; Briec & Lesourd, 1999).
References
Aparicio, J., Pastor, J. T., & Ray, S. C. (2013). An overall measure of technical inefficiency at the
firm and at the industry level: The ‘lost profit on outlay’. European Journal of Operational
Research, 226, 154–162.
Badunenko, O., Fritsch, M., & Stephan, A. (2008). Allocative efficiency measurement revisited—
Do we really need input prices? Economic Modelling, 25, 1093–1109.
References 63
Banker, R. D., Chang, H. H., & Cooper, W. W. (1996). Equivalence and implementation of
alternative methods for determining returns to scale in data envelopment analysis. European
Banker, R. D., Chang, H. H., & Natarajan, R. (2007). Estimating DEA technical and allocative
inefficiency using aggregate cost or revenue data. Journal of Productivity Analysis, 27,
115–121.
Banker, R. D., & Thrall, R. M. (1992). Estimation of returns to scale using data envelopment
Briec, W. (1997). A graph-type extension of Farrell technical efficiency measure. Journal of
Productivity Analysis, 8, 95–110.
Briec, W. (1999). Hölder distance function and measurement of technical efficiency. Journal of
Briec, W., & Lesourd, J. B. (1999). Metric distance function and profit: Some duality results.
Journal of Optimization Theory and Applications, 101, 15–33.
Chambers, R. G., Chung, Y., & Färe, R. (1996). Benefit and distance functions. Journal of
Economic Theory, 70, 407–418.
Chambers, R. G., Chung, Y., & Färe, R. (1998). Profit, directional distance functions, and
Nerlovian efficiency. Journal of Optimization Theory and Applications, 98, 351–364.
Chavas, J. P., & Cox, T. L. (1999). A generalized distance function and the analysis of production
efficiency. Southern Economic Journal, 66, 294–318.
Cherchye, L., Kuosmanen, T., & Post, T. (2001). FDH directional distance functions with an
application to European commercial banks. Journal of Productivity Analysis, 15, 201–215.
Deprins, D. L., Simar, L., & Tulkens, H. (1984). Measuring labor efficiency in post offices. In
M. Marchand, P. Pestieau, & H. Tulkens (Eds.), The performance of public enterprises:
Concepts and measurements (pp. 243–267). North-Holland.
Fukuyama, H., & Weber, W. L. (2003). Modeling input allocative efficiency via distance and quasi-
distance functions. Journal of the Operations Research Society of Japan, 46, 264–285.
Fukuyama, H., & Weber, W. L. (2005). Estimating output gains by means of Luenberger efficiency
measures. European Journal of Operational Research, 164, 535–547.
Kuosmanen, T. (2001). DEA with efficiency classification preserving conditional convexity.
Leleu, H. (2009). Mixing DEA and FDH models together. Journal of the Operational Research
Society, 60, 1730–1737.
Luenberger, D. G. (1992). Benefit functions and duality. Journal of Mathematical Economics, 21,
461–481.
Luenberger, D. G. (1995). Microeconomic theory. McGraw-Hill.
McFadden, D. (1978). Cost, revenue, and profit functions. In M. Fuss & D. McFadden (Eds.),
Production economics: A dual approach to theory and applications (Vol. 1, pp. 1–110). North-
Holland.
Nerlove, M. (1965). Estimation and Identification of Cobb-Douglas Production Functions. Rand
McNally.
Sahoo, B. K., Mehdiloozad, M., & Tone, K. (2014). Cost, revenue and profit efficiency measure-
ment in DEA: A directional distance function approach. European Journal of Operational
Research, 237, 921–931.
Shephard, R. W. (1953). Cost and production functions. Princeton University Press.
Soleimani-damaneh, M., Jahanshahloo, G. R., & Reshadi, M. (2006). On the estimation of returns
to scale in FDH models. European Journal of Operational Research, 174, 1055–1059.
Soleimani-damaneh, M., & Mostafaee, A. (2009). Stability of the classification of returns to scale in
FDH models. European Journal of Operational Research, 196, 1223–1228.
Sueyoshi, T. (1992). Measuring technical, allocative and overall efficiencies using a DEA algo-
rithm. Journal of the Operational Research Society, 43, 141–155.
Thrall, R. M. (1999). What is the economic meaning of FDH? Journal of Productivity Analysis, 11,
243–250.
Tulkens, H. (1993). On FDH efficiency analysis: Some methodological issues and application to
retail banking, courts, and urban transit. Journal of Productivity Analysis, 4, 183–210.
Zelenyuk, V. (2013). A scale elasticity measure for directional distance function and its dual:
Theory and DEA estimation. European Journal of Operational Research, 228, 592–600.
Zofio, J. L., Pastor, J. T., & Aparicio, J. (2013). The directional profit efficiency measure: On why
profit inefficiency is either technical or allocative. Journal of Productivity Analysis, 40,
257–266.
Chapter 4
Slacks-Based Efficiency Measures
The output–input ratio and radial distance function measures discussed in the
preceding two chapters are essentially the same approach expressed in different
forms, bearing the primal-dual relationship. The basic idea is to measure the relative
distance between the DMU being evaluated and its projection on the production
frontier along the ray from the origin pointing at the DMU. The efficiencies are thus
radial measures. In contrast, there are also non-radial measures, and slacks-based
measures are the major branch of these.
Radial measures can be discussed from either the input or output side, where the
former fixes the output at a specific level to measure efficiency, while the latter fixes
the input at a specific level. One weakness of this approach is that the efficiency
scores obtained for weakly efficient, or near weakly efficient, DMUs are not well
defined, and thus cannot be compared with inefficient DMUs, making the ranking
difficult. Moreover, since the input and output factors must be considered separately,
the results obtained from these two aspects may be inconsistent, which again makes
the rankings questionable. One approach to solve these problems is the use of slacks-
based measures.
Recall that the efficiency measured by the input model is essentially θ and is 1/φ
by the output model. The difference between the DMU being evaluated and its
benchmark is X0 - nj = 1 λj Xj , nj = 1 λj Y j - Y 0 , also called slack. The slacks-
based approach uses the slacks to measure performance. Since the measure is not
based on the ray from the origin along the DMU being evaluated, it is non-radial.
This approach comes in different forms, which have various names, although since
Tone (2001) the most common term is a slacks-based measure model, abbreviated as
an SBM model. In this chapter we will discuss three SBM models: the additive,
Russell measures, and Russell ratio versions. An example is then presented to show
the differences among the four major efficiency measurement models discussed in
Chaps. 2 and 3, as well as the current one.

https://doi.org/10.1007/978-3-031-27593-7_4
66 4 Slacks-Based Efficiency Measures
4.1 Additive Model
The first attempt to use slacks to measure performance was seen in the additive
model proposed by Charnes et al. (1985). Consider the general case of variable
returns to scale, with the production possibility set
n
T = ðX, Y Þ j = 1 λj X j ≤ X, -
n n
j = 1 λj Y j
≥ Y, j = 1 λj = 1, λj ≥ 0, j = 1, . . . , ng defined in Expression
(3.4). The idea is to maximize the sum of the slacks associated with the inputs,
si- , and outputs, sþ
r , in the production possibility set:
m s
max: si- þ sþ
r
i=1 r=1
n
s:t: λj X ij þ si- = X i0 , i = 1, . . . , m
j=1
n
ð4:1Þ
λj Y rj - sþ
r = Y r0 , r = 1, . . . , s
j=1
n
λj = 1
j=1
λj , si- , sþ
r ≥ 0, j = 1, . . . , n, i = 1, . . . , m, r = 1, . . . , s
At optimality, if all slack variables have a value of zero, then the DMU being
evaluated is efficient. If it is not the case, then the nonzero slacks show the amount of
the corresponding factor that can be improved to become efficient.
What Model (4.1) does is to search for an efficient point on the production frontier
that is farthest from the DMU being evaluated, in terms of 1-norm distance, as the
benchmark to measure the performance. The production frontier constructed via the
additive model (4.1) is the same as that constructed via the BCC input or output
models. Different from the BCC model that has two forms, input and output, the
additive model has only one, which is a combination of input and output. To see how
this idea works, consider six DMUs using one-input X to produce one-output Y, with
the data shown in columns two and three of Table 4.1. By applying Model (4.1), one
obtains an optimal objective value of zero for DMUs A, B, C, and D, indicating that
they are efficient, and the production frontier is composed of the piecewise line
segments ABCD, as shown in Fig. 4.1. DMUs E and F are inefficient, in that the
former has s- = 4, s+ = 1, λB = 1, with DMU B as the benchmark, and the latter
has s- = 4.5, λB = λC = 0:5, with U = (5.5, 5)as the benchmark.
The 1-norm distance for DMU E is s- + s+ = 5, which is greater than that of
DMU F, 4.5. However, this does not imply that DMU F outperforms DMU E,
because this approach is not unit invariant, different from the radial measure. Recall
that the envelopment form of the BCC model has the constraints of nj = 1 λj X ij ≤
4.1 Additive Model 67
Table 4.1 An example of performance measurement using the additive model

DMU Input X Output Y Slacks λj Benchmark
A 3 2 0 λA = 1 A
B 4 4 0 λB = 1 B
C 7 6 0 λC = 1 C
D 10 7 0 λD = 1 D
E 8 3 s- = 4, s+ = 1 λB = 1 B
F 10 5 s- = 4.5 λB = λC = 0:5 U = (5.5, 5)
Fig. 4.1 Additive model for Y

measuring performance
10
8
D
C
6
U F
–
B s
4 X+2–1.5Y = 0 +
s
– E
s
2 A
O 2 4 6 8 10 X
θX i0 and nj= 1 λj Y rj ≥ Y r0 , with the objective function of “min. θ”. The efficiency
score θ will not be altered by using different units, e.g., centimeters or kilometers, to
measure a factor, because the factor c in (cXij, cXi0) and (cYrj, cYr0) in the constraints
will cancel out, thus producing the same value for θ. In the multiplier form of the
s m
BCC model the constraints are r = 1 ur Y rj - i = 1 vi X ij - v0 ≤ 0, so the effects
caused by changing the units of measurement in Yr and Xi can be absorbed by the
corresponding multipliers ur and vi, similar to that in the regression analysis. The
BCC (or CCR) model is thus units invariant. In contrast, the objective function of the
- þ
additive model is m i = 1 si þ
s
r = 1 sr . When the unit of a factor is changed c times,
the corresponding constraint becomes nj = 1 λj cX ij þ si- = ðcX i0 Þ, which implies
that si- = csi- . In other words, the corresponding slack variable will also be changed
c times, to result in a different objective value. For the example in Table 4.1, suppose
input X was measured in inches. If the unit of X is changed to centimeters, and the
unit of Y remains the same, then the objective value for DMU E becomes
2.54 × 4 + 1 = 11.16, and that of DMU F becomes 2.54 × 4.5 + 0 = 11.43, making
the ranks of E and F reversed.
Although the additive model is not unit invariant, it is translation invariant (Ali &
Seiford, 1990), in that adding a constant to all observations of a factor does not alter
the solution. For example, when all observations of input i are increased by c, the
n -
corresponding constraint becomes j = 1 λj X ij þ c þ si = ðX i0 þ cÞ, or
n n - n
j=1 λ j X ij þ c j=1 λ j þ s i = X i0 þ c: Since j=1 λ j = 1, the constant c on both
sides of the equation cancels out to maintain the constraint of the original form. This
property is desirable when one has negative observations for some DMUs, which
cause difficulties in measuring efficiencies, as an appropriate number c can then be
added to all DMUs to make all observations positive. Note that this property does not
hold under constant returns to scale, as nj = 1 λj in this case is not guaranteed to equal
one. This property is obviously not possessed by the BCC (or CCR) model, because
n
the constant c cannot be eliminated in the constraint of j = 1 λj X ij þ c ≤
θðX i0 þ cÞ:
To gain the property of unit invariance, Cooper et al. (1999) proposed two
variants of this model, namely the Measure of Inefficiency Proportions (MIP) and
Range-Adjusted Measure (RAM) of inefficiency. The MIP modifies the objective
function by dividing the slack variables by their corresponding DMU observations,
as follows:
m
si- s
sþ
F 0 = max: þ r
ð4:2Þ
i=1
X i0 r = 1 Y r0
With this modification the units being changed to the slack variables will cancel
out with those being changed to the observations, making the relative deviations
from the benchmark for all factors comparable. An efficient DMU has an F0 value of
zero. Larger values of si- and sþ r imply that the corresponding DMU is less efficient.
F0 thus serves as an index of inefficiency. For all these models, if the returns to scale
are assumed to be constant, then one simply deletes the convexity constraint of
n
j = 1 λj = 1 from the constraint set.
Similar to the MIP, the RAM modifies the objective function by dividing the
slack variables by the ranges of the corresponding factors, as follows:
1
m
si- s
sþ
F 0 = max: - þ
r
þ ð4:3Þ
mþs i=1
Ri R
r=1 r
where Ri- and Rþ

r are the ranges of input i and output r, respectively, defined as:
4.2 Russell Measures 69
Ri- = max X ij - min X ij

j j
ð4:4Þ
Rþ
r = max Y rj - min Y rj
j j
If a DMU is efficient, then all slack variables are zero, with F0 = 0. It is shown in
Cooper et al. (1999) that 0 ≤ F0 ≤ 1. Thus, 1 - F0 serves as an index for efficiency.
m s
i=1 r=1
m s
s:t: vi X ij þ v0 - ur Y rj ≥ 0, j = 1, . . . , n ð4:5Þ
i=1 r=1
ur , vi ≥ 1, r = 1, . . . , s, i = 1, . . . , m
The constraints restrict the choice of the multipliers to those greater than or equal
to one, and the resulting virtual input is greater than or equal to the resulting virtual
output. The objective is to make the virtual input and output of the DMU being
evaluated as close as possible, so that the efficiency is maximal. Geometrically, the
DMU being evaluated is searching for the supporting hyperplane
m s
i = 1 vi X i þ v0 - r = 1 ur Y r = 0, where Xi and Yr are variables, at the benchmark
of the focal DMU. For example, DMU F has an optimal solution of v1 = 1, v0 = 2,
u = 1.5, indicating that the corresponding supporting hyperplane at its benchmark
U is X + 2 – 1.5Y = 0, as shown in Fig. 4.1. Obviously, the supporting hyperplane is
not unique, as is clear for the case of DMU B.
4.2 Russell Measures
In radial measures, although a weakly efficient DMU can be identified with the help
of the non-Archimedean number ε, how to determine a suitable efficiency score for
this DMU is still a problem. To solve this, Färe and Lovell (1978) generalized the
radial measures to the Russell measures of efficiency. There are three forms: input,
output, and input–output average.
4.2.1 Input model
The Shephard’s input distance function, referring to Model (3.7), is defined as Φ(Y,
X) = 1/θ(Y, X), where θ(Y, X) = min . {θ | θX 2 Q(Y), θ ≥ 0}. The Russell’s input
measure of efficiency is used to allow each input factor i to have a different
contraction parameter θi, and generalize the input distance function as θðY, XÞ =
m
min: i = 1 θi =m ðθ1 X 1 , θ2 X 2 , . . . , θm X m Þ 2 QðY Þ, θi 2 ð0, 1, i = 1, . . . , m :
The Russell input efficiency of a DMU under constant returns to scale is thus
measured as:
m
1
R0 = min: θi
m i=1
n
s:t: λj X ij ≤ θi X i0 , i = 1, . . . , m
j=1
n
ð4:6Þ
λj Y rj ≥ Y r0 , r = 1, . . . , s
j=1
θ ≤ 1, i = 1, . . . , m
λj ≥ 0, j = 1, . . . , n
Conceptually, a benchmark on the frontier is sought, such that the average of the
contraction factors of all inputs measured from this benchmark is minimal. Let X be
the benchmark, then θi = X i =X i0 is the efficiency with respect to the ith input, and
m
i = 1 θi =m is the average efficiency of all inputs. The constraints of θi ≤ 1 restrict the
benchmark selected for measuring efficiency to the dominating points on the fron-
tier. As discussed in Chap. 3, θi is the contraction factor that maintains θiXi0 in the
production possibility set. In practice, unless Xi0 is equal to zero, θi will definitely be
positive. The lower bound constraint of θi > 0 for θi is thus not needed to express the
condition of θi 2 (0, 1].
Consider the example in Table 4.2, copied from Table 3.2, where five DMUs
apply different amounts of inputs X1 and X2 to produce one unit of output Y. Column
five shows the efficiency measured from the CCR model. The piecewise line
segments ABC in Fig. 4.2 constitute the isoquant, where DMUs A, B, and C on the
isoquant are efficient. By applying Model (4.6), the efficiency of each DMU is
calculated, with the optimal nonzero variables shown in the sixth to ninth columns of
Table 4.2. Efficient DMUs have the benchmarks of themselves. For the inefficient
DMU D the benchmark is DMU B, as indicated by the only nonzero linear combi-
nation variable λB = 1, and the associated θi are θ1 = X 1B =X 1D = 0:5 and θ2 =
X 2B =X 2D = 2=3 to result in RD = θ1 þ θ2 =2 = 7=12: Similarly, the inefficient
Table 4.2 An example of efficiency measurement using the Russell input model
DMU Input X1 Input X2 Output Y E0 θ1 θ2 R0 λj Benchmark
A 1 4 1 1 1 1 1 λA = 1 A
B 2 2 1 1 1 1 1 λB = 1 B
C 4 1 1 1 1 1 1 λC = 1 C
D 4 3 1 0.6 0.5 2/3 7/12 λB = 1 B
E 6 1 1 1–2ε 2/3 1 5/6 λC = 1 C
Fig. 4.2 The Russell input

model for measuring X2
efficiency
A
4
q1=X1B/X1D
D
q2=X2B/X2D
2 B q1=X1C/X1E
q2=X2C/X2E
C E
O 2 4 6 X1
DMU E selects DMU C as the benchmark, to result in θ1 = X 1C =X 1E = 2=3 and

θ2 = X 2C =X 2E = 1, and RE = (2/3 + 1)/2 = 5/6. This DMU is weakly efficient, with
ED = 1 - 2ε measured from the CCR model. As mentioned above, the efficiency
score 1 - 2ε does not properly show the performance of this DMU. The Russell
measure of 5/6, in contrast, is more representative for the efficiency of this DMU.
The Russell input model (4.6) is a generalization of the CCR input model, which
boils down to the latter when all parameters θi are required to be the same. Since the
CCR model is more restrictive, its efficiency measure is greater than or equal to the
Russell efficiency measure, and this is verified by comparing E0 with R0 contained in
Table 4.2. This feature can also be discussed from the dual formulation. To make the
explanation clearer, consider a simpler form of Model (4.6), where the constraints
θi ≤ 1, i = 1, . . ., m, are ignored. The dual is:
s
max: ur Y r0
r=1
s:t: vi X i0 = 1=m, i = 1, . . . , m
s m
ð4:7Þ
ur Y rj - vi X ij ≤ 0, j = 1, . . . , n
r=1 i=1
ur , vi ≥ 0, r = 1, . . . , s, i = 1, . . . , m
When the m constraints of viXi0 = 1/m, i = 1, . . ., m, are replaced with their sum
of m i = 1 vi X i0 = 1, the CCR model (2.3) is obtained. Since the constraints of Model
(4.7) are more stringent than those of the CCR model, the efficiency calculated from
the former is less than or equal to that calculated from the latter.
Model (4.6) measures efficiency based on the performances of individual inputs.
From the above example it is noted that the contraction factor of each input, θi, is
related to the slack between the DMU being evaluated and its benchmark. More
specifically, if we set θi = X i0 - si- =X i0 = 1 - si- =X i0 , then Model (4.6)
becomes:
1
m
si-
R0 = min: 1 -
m X
i = 1 i0
n
s:t: λj X ij = X i0 - si- , i = 1, . . . , m
j=1 ð4:8Þ
n
λj Y rj ≥ Y r0 , r = 1, . . . , s
j=1
λj , si- ≥ 0, j = 1, . . . , n, i = 1, . . . , m
which is equivalent to the MIP model (4.2), with the effects of output slacks ignored.
The Russell input measure is thus also a slacks-based measure. Note that the first set
of constraints is expressed in equality form, instead of less-than-or-equal-to form.
The reason is because - si- is to be minimized, which will be achieved only when
both sides of the constraint are equal.
4.2.2 Output Model
The Russell measure can also be defined from the output side by allowing each
output to have different expansion factors φr (Färe et al., 1983). Under constant
returns to scale the model is:
s
1 1
= max: φr
R0 s r=1
n
s:t: λj X ij ≤ X i0 , i = 1, . . . , m
j=1
n
ð4:9Þ
λj Y rj ≥ φr Y r0 , r = 1, . . . , s
j=1
φr ≥ 1, r = 1, . . . , s
λj ≥ 0 j = 1, . . . , n
The constraints φr ≥ 1, r = 1, . . ., s require the DMU to use a dominating point on

the frontier as the benchmark to measure efficiency. For cases of variable returns to
scale, one simply adds the convexity constraint nj = 1 λj = 1 to Model (4.9). If all φr
variables are required to be the same, then Model (4.9) boils down to the envelop-
ment form of the CCR output model (2.7), ignoring the non-Archimedean number ε.
Since Model (4.9) is more flexible than Model (2.7), the objective value of the
Table 4.3 An example of efficiency measurement using the Russell output model
DMU X Y1 Y2 E0 φ1 φ2 R0 λj Benchmark
A 1 2 3 2/3 2 4/3 3/5 λD = 1 D
B 1 2 5 1 1 1 1 λB = 1 B
C 1 4 1 1/(1.25 + 0.75ε) 1 4 2/5 λD = 1 D
D 1 4 4 1 1 1 1 λD = 1 D
E 1 5 2 1 1 1 1 λE = 1 E
former is greater than or equal to that of the latter, which implies that the efficiency
calculated from the former, 1= sr = 1 φr =s , is less than or equal to that calculated
from the latter, E0.
To see how this model works, consider the data in Table 4.3, which is copied
from Table 3.4, where five DMUs apply one unit of input X to produce different
amounts of outputs Y1 and Y2. E0 in column five is the efficiency measured from the
CCR model. The optimal solutions for each DMU obtained from Model (4.9) are
shown in the sixth to ninth columns of Table 4.3. Of the five DMUs, B, D, and E are
efficient. The product transformation curve, as shown in Fig. 4.3, is composed of the
line segments BDE. Both of the two inefficient DMUs, A and C, use DMU D as the
benchmark to measure efficiency, as indicated by the nonzero linear combination
variable of λD = 1. The φr value is the ratio of the rth coordinate of the benchmark to
that of the DMU being evaluated, which is φ1 = Y 1D =Y 1A = 2 and φ2 =
Y 2D =Y 2A = 4=3 for DMU A, and φ1 = Y 1D =Y 1C = 1 and φ2 = Y 2D =Y 2C = 4 for
DMU C, with RA = 2= φ1 þ φ2 = 3=5 and RC = 2= φ1 þ φ2 = 2=5, respectively.
As expected, the output efficiency scores measured from the Russell model, R0, are
less than or equal to those measured from the CCR model, E0.
Similar to the input case of Sect. 4.2.1, if the φr ≥ 1 constraints of Model (4.9) are
ignored for simplicity of discussion, then its dual can be formulated as:
m
min: vi X i0
i=1
s:t: ur Y r0 = 1=s, r = 1, . . . , s
m s
ð4:10Þ
vi X ij - ur Y rj ≥ 0, j = 1, . . . , n
i=1 r=1
ur , vi ≥ 0, r = 1, . . . , s, i = 1, . . . , m
Adding the s constraints of urYr0 = 1/s, r = 1, . . ., s together, one obtains

s
= 1, making it the same as the CCR output model (2.6). The objective
r = 1 ur Y r0
value of Model (4.10) thus is greater than or equal to that of the CCR model (2.6),
indicating that the Russell output efficiency is less than or equal to that of the CCR
output efficiency.
Fig. 4.3 The Russell output Y2

model for measuring
efficiency
6
B
j =Y1D/Y1A
4 D
j2=Y2D/Y2A
A
2 E
j 2=Y2D/Y2C
C
j1=Y1D/Y1C
O 2 4 6 Y1
Model (4.9) can also be converted to a slacks-based model, as seen in Sect. 4.2.1
from the input side. Let φr = Y r0 þ sþ þ
r =Y r0 = 1 þ sr =Y r0 , then Model (4.9)
becomes:
1 1
s
sþ
= max: 1 þ r
R0 s Y
r = 1 r0
n
s:t: λj X ij ≤ X i0 , i = 1, . . . , m
j=1 ð4:11Þ
n
λj Y rj = Y r0 þ sþ
r , r = 1, . . . , s
j=1
λj , sþ
r ≥ 0, j = 1, . . . , n, r = 1, . . . , s
which is equivalent to the MIP model (4.2), without considering the input slacks.
The Russell output efficiency measure is thus also a slacks-based measure. Similar to
the case of Model (4.8), the second set of constraints is expressed in equality form, as
a convention.
4.2.3 Input–Output Average Model
The previous two subsections discuss models from the input and output sides
separately as a generalization of the conventional CCR input and output models.
Färe et al. (1985) combined them together to propose the following Russell Graph
Measure of Technical Efficiency:
m s
1 1
R0 = min: θi þ
mþs i=1
φ
r=1 r
n
s:t: λj X ij ≤ θi X i0 , i = 1, . . . , m
j=1
n ð4:12Þ
λj Y rj ≥ φr Y ro , r = 1, . . . , s
j=1
θi ≤ 1, i = 1, . . . , m
φr ≥ 1, r = 1, . . . , s
λj ≥ 0, j = 1, . . . , n
As explained in the input and output measures, the constraints of θi ≤ 1 and

φr ≥ 1 are imposed to restrict the benchmark to the points on the frontier that
dominate the DMU being evaluated. The convexity constraint of nj = 1 λj = 1 can be
included if the returns to scale are assumed to be variable.
The constraints of Model (4.12) use the contraction factors θi for the inputs and
expansion factors φr for the outputs to express the production possibility set. The
objective function is the average of the contraction and expansion factors of the
m individual inputs and s individual outputs (in reciprocal form). Since this model
takes both the input and output into consideration in measuring efficiencies, it is
more appropriate. The only inconvenience is that the model is nonlinear, and thus
not many problem solvers are available, meaning that it is not easy to obtain the
solutions.
In order to get some idea of how Model (4.12) measures efficiencies, consider the
data in Table 4.4, copied from Table 3.1, where five DMUs applying different
amounts of input X to produce different amounts of output Y. By applying Model
(4.12), the optimal θ, φ, R0, and nonzero λj for each DMU are obtained, as shown
in columns four to seven of Table 4.4, under the heading of “Constant returns to
scale.” The optimal values for λj also show the benchmark (abbreviated as BM in
p
Table 4.4) of the corresponding DMU. For example, λj for DMU A is λD = 1= 6,

p p that the benchmark for DMU A is R = λD ðX D , Y D Þ = ð2:4495, 2:4495Þ, or
indicating
6, 6 to be exact, as shown in Fig. 4.4. The optimal contraction factor for the
input is θ = X R =X A = 2=3, the optimal expansion factor for the output is
φ = Y R =Y A = 3=2, and the efficiency is R0 = ðθ þ 1=φ Þ=2 = 2=3. DMUs
B and D are efficient, with θ = φ = 1, and the benchmarks are themselves. DMUs
C and E have θ = 1=φ = 3=5 and 4=5, respectively. They have the same

reference
p pointp D, with λD = p 5=12 p and 20=9, and different benchmarks of
S = 15, 15 , and T = 4 5, 4 5 .
Under variable returns to scale, the convexity constraint nj = 1 λj = 1 is added to
Model (4.12) to measure efficiency. Of the five DMUs, only DMU C is inefficient,
with the benchmark of DMU B. The optimal θ* and φ* are equal to XB/XC = 0.8 and
76
Table 4.4 An example of efficiency measurement using the Russell input–output average model
Constant returns to scale Variable returns to scale
DMU X Y θ* φ* R0 λj BM θ* φ* R0 λj BM
p
A 3 2 2=3 3=2 2=3 λD = 1= 6 R 1 1 1 λA = 1 A
B 4 4 1 1 1 λB = 1 B 1 1 1 λB = 1 B
C 5 3 3=5 5=3 3=5 λD = 5=12 S 0.8 4/3 0.775 λB = 1 B
D 6 6 1 1 1 λD =1 D 1 1 1 λD = 1 D
4
E 10 8 4=5 5=4 4=5 λD = 20=9 T 1 1 1 λE = 1 E

Slacks-Based Efficiency Measures
Fig. 4.4 The Russell input– Y

output average model for
measuring efficiency
10
T
8
E
6
D
V
B
4 S
U C
R W
2 A
O 2 4 6 8 10 X
YB/YC = 4/3, respectively, with an average efficiency of R0 = (0.8 + 3/4)/2 = 0.775.

The last five columns of Table 4.4, under the heading of “Variable returns to scale”,
show the nonzero optimal variables and the benchmarks of the five DMUs.
Recall that in Sect. 3.2 the benchmark for DMU C measured by the CCR input
model was point U (in Fig. 4.4) and was point V when measured by the CCR output
model. Here, when the effects of both input and output are considered, the bench-
mark becomes a compromise point S, located between points U and V on the frontier
OT. Note that the Russell input and output measures coincide with the radial input
and output measures, respectively, when there is only one input and one output. The
efficiency score measured from the input–output average model, which is 3=5 ffi
0:7746 in this example, is greater than the input and output efficiency of 0.6. In Sect.
3.3 the benchmarks for DMU C obtained from the BCC input and output models
were points W and V, respectively, with efficiency scores of 0.7 and 0.6, which are
also less than the 0.775 obtained with the Russell input–output average model, as
expected.
Since Model (4.12) is a combination of Models (4.6) and (4.9), it can also be
converted to a slacks-based model by substituting θi = X i0 - si- =X i0 and φr =
Y r0 þ sþ
r =Y r0 :
1 X i0 - si-
m s
Y r0
R0 = min: þ
mþs i=1
X i0 r=1
Y r0 þ sþ
r
n
s:t: λj X ij = X i0 - si- , i = 1, . . . , m
j=1
n
λj Y rj = Y ro þ sþ
r , r = 1, . . . , s
j=1
λj , si- , sþ
r ≥ 0, j = 1, . . . , n, i = 1, . . . , m, r = 1, . . . , s
ð4:13Þ
The largest value for R0 is one, when all slack variables si- and sþ
r are zero.
This conversion shows that the Russell measures, including input, output, and
input–output average, are slacks-based measures.
4.3 Russell Ratio Model
The Russell input–output average model (4.12) takes both the input and output into
consideration in measuring efficiency and is thus a more appropriate one. However,
it is nonlinear, and so not many problem solvers are available to obtain a solution.
The objective function of Model (4.12) is the average of the contraction and
expansion factors of the m inputs and s outputs. Another approach is to take the
ratio of the average contraction factor of the m inputs and the average expansion
factor of the s outputs to form the following Russell input–output ratio model:
1 m
θ
i=1 i
R0 = min: m
1 s
φ
r=1 r
s
n
s:t: λj X ij ≤ θi X i0 , i = 1, . . . , m
j=1
n
ð4:14Þ
λj Y rj ≥ φr Y r0 , r = 1, . . . , s
j=1
θi ≤ 1, i = 1, . . . , m
φr ≥ 1, r = 1, . . . , s
λj ≥ 0, j = 1, . . . , n
The objective function is the product of the Russell input arithmetic average
efficiency and output harmonic average efficiency. This model was proposed by
Pastor et al. (1999) and is called the Enhanced Russell Graph Efficiency Measure.
4.3 Russell Ratio Model 79
Model (4.14) can be converted to a slacks-based measure by substituting θi =

1 - si- =X i0 and φr = 1 þ sþ
r =Y r0 , as carried out in Sect. 4.2.3, to obtain:
1 m
1- s - =X i0
i=1 i
R0 = min: m
1 s
1þ sþ =Y r0
r=1 r
s
n
s:t: λj X ij þ si- = X i0 , i = 1, . . . , m
j=1
n
λj Y rj - sþ
r = Y r0 , r = 1, . . . , s
j=1
λj , si- , sþ
r ≥ 0, j = 1, . . . , n, i = 1, . . . , m, r = 1, . . . , s
ð4:15Þ
This efficiency measure was termed slacks-based measure (SBM) by

Tone (2001).
Model (4.15) is nonlinear; however, it can be linearized by applying the idea in
Charnes and Cooper (1962) for transforming linear fractionals into linear functions.
The idea is to set the denominator of the objective function to the inverse of a
variable t, t = 1= 1 þ 1s sr = 1 sþ
r =Y r0 , to be a constraint, and replace the objective
-
function to t 1 - m1 m s
i=1 i =X i0 , to obtain the following linear program:
1
m
si-
R0 = min: t -
m X
i = 1 i0
1
s
sþ
s:t: tþ r
=1
s Y
r = 1 r0
n
λj X ij þ si- = tX i0 , i = 1, . . . , m
j=1
n
λj Y rj - sþ
r = tY r0 , r = 1, . . . , s
j=1
λj , si- , sþ
r ≥ 0, j = 1, . . . , n, i = 1, . . . , m, r = 1, . . . , s
ð4:16Þ
where si- = tsi- , sþ þ

r = tsr , and λj = tλj : The variable t must be strictly positive.
-
Since λj and si are positive, t will be automatically positive from the second set of
constraints, and the constraint of t > 0 is thus not needed. If the returns to scale are
assumed to be variable, then the constraint nj = 1 λj = t is added. Note that this
constraint is the convexity constraint of nj = 1 λj = 1, with λj being replaced with
λj =t.
Recall that the CCR input model has the following form:
m s
E0 = min: θ-ε t i- þ tþ
r
i=1 r=1
n
s:t: λj X ij þ t i- = θX, i0 i = 1, . . . , m
j=1
n
λj Y rj - t þ
r = Y r0 , r = 1, . . . , s
j=1
λj , t i- , t þ
r ≥ 0, j = 1, . . . , n, i = 1, . . . , m, r = 1, . . . , s
ð4:17Þ
Let θ , t i- , t þ
r be an optimal solution to this model, then
θ X i0 - t i- , Y r0 þ t þ r corresponds to a projection point
n n
j = 1 λj X ij , j = 1 λj Y rj on the frontier, implying that λ* is feasible to the SBM
model (4.15). In other words, X i0 - si- , Y r0 þ sþ
r in Model (4.15) can be replaced
with θ X i0 - t i , Y r0 þ t r to obtain the relationships of θ X i0 - t i- = X i0 - si-
- þ

and Y r0 þ t þ þ -
r = Y r0 þ sr , or si = ð1 - θ ÞX i0 þ t i
-
and sþ þ
r = t r : Substituting
these into the objective function of Model (4.15), one obtains the following
relationship:
1- 1 m
i=1 ð1 - θ ÞX i0 þ t i- =X i0 θ - m1 m -
i = 1 t i =X i0
R0 = m
= ≤ θ ð4:18Þ
1 þ 1s sr = 1 t þ
r =Y r0 1 þ 1s s þ
r = 1 t r =Y r0
In other words, the SBM efficiency R0 is less than or equal to the CCR input
efficiency θ*. This relationship also applies to the BCC input efficiency, where the
returns to scale are assumed to be variable.
Similarly, the CCR output model has the following form:
m s
1
= max: φþε t i- þ tþ
r
E0 i=1 r=1
n
s:t: λj X ij þ t i- = X i0 , i = 1, . . . , m
j=1
n
λj Y rj - t þ
r = φY r0 , r = 1, . . . , s
j=1
λj , t i- , t þ
r ≥ 0, j = 1, . . . , n, i = 1, . . . , m, r = 1, . . . , s
ð4:19Þ
4.3 Russell Ratio Model 81
At optimality, X i0 - t i- , φ Y r0 þ t þ
r corresponds to a projection point
n n
j = 1 λj X ij , j = 1 λj Y rj on the frontier, implying that λ is feasible to the SBM
*
model (4.15). That is, X i0 - si- , Y r0 þ sþ r in Model (4.15) can be replaced with
X i0 - t i- , φ Y r0 þ t þ r to obtain the relationships of si- = t i- and
sþ
r = ð φ - 1 Þ Y r0 þ t þ
r . Substituting these into the objective function of Model
(4.15), one obtains:
m - m -
1- 1
i = 1 t i =X i0 1 - m1 i = 1 t i =X i0 1
R0 = m
= ≤ ð4:20Þ
1 þ 1s s
r=1 ðφ - 1ÞY r0 þ t þ
r =Y r0 φ þ 1s s
r = 1 t r =Y r0
þ φ
The SBM efficiency is thus less than or equal to the CCR output efficiency, and
this property also holds under variable returns to scale.
The linearized SBM model (4.16) has a dual, which, under variable returns to
scale, with the constraint of nj = 1 λj = t included, can be formulated as:
max: η
m s
s:t: vi X i0 þ v0 - ur Y r0 þ η = 1
i=1 r=1
s m
ur Y rj - vi X ij - v0 ≤ 0, j = 1, . . . , n ð4:21Þ
r=1 i=1
vi ≥ 1=ðmX i0 Þ, i = 1, . . . , m
ur ≥ η=ðsY r0 Þ, r = 1, . . . , s
Replacing η in the objective function and the bound constraints with the rela-
tionship of the first constraint, the following model is obtained:
s m
max: 1 þ ur Y r0 - vi X i0 - v0
r=1 i=1
m s
s:t: vi X ij þ v0 - ur Y rj ≥ 0, j = 1, . . . , n
i=1 r=1
ð4:22Þ
1
vi ≥ , i = 1, . . . , m
mX i0
s m
1þ -
r = 1 ur Y r0 i = 1 vi X i0 - v0
ur ≥ , r = 1, . . . , s
sY r0
The constant 1 in the objective function can be omitted without altering the
optimal solution. However, it must be included to obtain the correct efficiency
score for the DMU being evaluated. At optimality, the equation sr = 1 ur Y r -

m
- v0 = 0 in the objective function is the supporting hyperplane at the
i = 1 vi X i
benchmark that the focal DMU uses to measure efficiency.
Consider the example in Table 4.1, where six DMUs apply different amounts of
input X to produce different amounts of output Y. The production frontier under
variable returns to scale is composed of the line segments ABCD, depicted in
Fig. 4.1, where DMUs A, B, C, and D are efficient. By applying Model (4.16),
with the variable returns to scale constraint of nj = 1 λj = t included, one obtains a
perfect efficiency of one for the four efficient DMUs. The optimal solutions for the

two inefficient DMUs, E and F, are R0 , t , si- , sþ
r , λj =

0:375, 0:75, 3, 0:75, λB = 0:75 for DMU E and 0:55, 1, 4:5, 0, λB = λC = 0:5
for DMU F, indicating that the benchmarks are (8 - 3/0.75, 3 + 0.75/
0.75) = (4, 4) = B for DMU E and (10 - 4.5, 5 + 0) = (5.5, 5) = U for DMU F.
The benchmarks are the same as those obtained from the additive model discussed in
Sect. 4.1.
When the dual model (4.22) is used to measure efficiency, the four DMUs on the
frontier, A, B, C, and D, also have an efficiency score of one, and the two inefficient
DMUs, E and F, have efficiency scores of 0.375 and 0.55, respectively. The optimal
multipliers for DMU F are u = 0.15, v = 0.1, and v0 = 0:2, indicating that the
supporting hyperplane at the benchmark U is 0.15Y - 0.1X - 0.2 = 0, or X + 2 -
1.5Y = 0, which is the same as that obtained from the additive model. The optimal
multipliers for DMU E are v = u = 0.125, and v0 = 0: The supporting hyperplane
for its benchmark, DMU B, is thus 0.125 Y - 0.125X = 0, or Y - X = 0, which is the
line passing through the origin and DMU B. From Fig. 4.1 it is clear that the feasible
supporting hyperplane at DMU B is not unique, and all the lines passing through
DMU B with a slope greater than or equal to that of the line passing through points
B and C and less than or equal to that of the line passing through points B and A are
legitimate ones.
4.4 A Classification of Efficiency Measures
This chapter and the preceding two discuss the major DEA models for measuring
efficiencies. Basically, the measures can be classified as radial and non-radial, as
shown in Fig. 4.5. The radial measure uses the ray emanating from the origin passing
through the DMU being evaluated to measure efficiencies. It has two forms: input-
and output-oriented. The output–input ratio models discussed in Chap. 2 and the
Shephard’s distance functions discussed in Chap. 3 belong to this category. The
output–input ratio and distance function models have a primal-dual relationship to
yield the same results and can thus be considered as the same model. However, they
provide different information, where the ratio models show the supporting hyper-
planes while the distance function models show the benchmarks. The merit of the
radial measure is that the efficiencies measured from individual inputs or outputs are
4.4 A Classification of Efficiency Measures 83
Fig. 4.5 A classification of

efficiency measures Input
Radial
Output
Efficiency
Measures Equal-proportional
Non-radial
Non-proportional
the same, making the subsequent task of aggregating factor efficiencies to form the
overall efficiency unnecessary. Its weaknesses are that only one category of input
and output is considered in measuring efficiencies, weakly efficient DMUs do not
have well-defined efficiency scores, and the efficiencies measured from the input and
output sides may be inconsistent.
The non-radial measures can be further classified as equal-proportional and
non-proportional. The equal-proportional measure requires all input and output
factors to have the same proportional measures. The directional distance function
measures discussed in Sect. 3.4 of Chap. 3 belong to this category. Their merit is that
both the input and output are taken into account when measuring performance.
Moreover, the contraction and expansion parameters measured for all input and
output factors are the same, resulting in a consistent overall performance measure.
Their weakness is the same as that of the radial measures, in that weakly efficient
DMUs do not have well-defined efficiency scores to be compared with other
inefficient ones.
The non-proportional measures allow all individual inputs and outputs to have
different efficiency measures, and the slacks-based measures discussed in this
chapter are representative ones of this category. This type of measure has several
merits, as it takes both the input and output into account in measuring efficiency and
provides well-defined efficiency scores for weakly efficient DMUs. The weakness is
that a set of persuasive weights is required when the performances of all input and
output factors are aggregated to form the overall measure, and this is usually lacking.
Another weakness, which often occurs in practice, is that when one of the coordi-
nates of the benchmark is relatively far away from that of the DMU being evaluated,
the corresponding slack variable will be very large, thus resulting in an unreasonably
small efficiency measure, making the rankings questionable. An example is
presented below to illustrate the features of several important DEA models for
measuring efficiencies.
Consider seven DMUs using different amounts of input X to produce different
amounts of output Y, with the data shown in columns two and three of Table 4.5. The
line segments ACDF in Fig. 4.6 constitute the production frontier, where DMUs
A, C, D, and F on the frontier are efficient. By applying the BCC input model (2.4)
and output model (2.7), the efficiencies of the seven DMUs are calculated, as shown
in the fourth and fifth columns of Table 4.5, respectively. The three inefficient
DMUs, B, E, and G are ranked fifth, seventh, and sixth, respectively, by the input
Table 4.5 An example comparing some models for efficiency measurement

Russell
BCC Directional dist. Russell average ratio
DMU X Y Input Output (-X0, Y0), η* θ* 1/φ* R0 (SBM)
A 3 2 1 1 0 1 1 1 1
B 4 0.01 0.75- 0.0025 0.25 + 1.9875ε 0.75 1/200 0.3775 0.0025
1.99ε (7) (6) (7) (7)
(5)
C 4 4 1 1 0 1 1 1 1
D 6 6 1 1 0 1 1 1 1
p p p
E 10 2 0.3 (7) 1/4 (6) 7/11 (7) 1= 5 1= 5 1= 5 0.2 (6)
(6)
F 10 8 1 1 0 1 1 1 1
G 15 8 2/3 (6) 1/ 5ε (5) 2/3 1 5/6 (5) 2/3 (5)
(1 + 5ε)
(5)
model and are ranked seventh, sixth, and fifth by the output model. The rankings for
the inefficient DMUs (numbers in parentheses in Table 4.5) are totally different.
Notice the unexpectedly high efficiency score for the weakly efficient DMU G and
the very small score for DMU B measured from the output side.
The directional distance function measure, along the direction of (-X0, Y0) based
on Model (3.20), is not able to give a suitable score to the weakly efficient DMU G,
either, although it takes both the input and output into account in measuring
efficiencies. The benchmarks for the three inefficient DMUs are points
A, K = (40/11, 36/11), and F (referring to Fig. 4.6). The distance parameter η
shows the closeness of the DMU to the frontier, and smaller values are considered
more efficient. The results show that the weakly efficient DMU G is ranked fifth, and
DMUs B and E are ranked sixth and seventh, respectively.
The values for θ, 1/φ, and R0 obtained from the Russell average model (4.12) are
shown in columns seven to nine of Table 4.5. Averaging the values of θ and 1/φ
eliminates extremely small scores measured by either the input or output model and
yields a reasonable overall score. The ranks for DMUs B,p E, and
pG are seventh, sixth,
and fifth, respectively. Their benchmarks are A, R = 2 5, 2 5 , and F, with the
slacks subscripted with “R” in Fig. 4.6.
Finally, the Russell ratio, or the SBM, Model (4.16), yields efficiency scores of
0.0025, 0.2, and 2/3 for DMUs B, E, and G, respectively. Their benchmarks are
DMUs C, D, and F, with the slacks subscripted with “SBM” in Fig. 4.6. Note the
very small efficiency score of 0.0025 for DMU B, which is the same as that measured
from the BCC output model. Since the SBM model uses the product of the Russell
input and output efficiencies as the efficiency measure, an extremely small effi-
ciency, such as that of DMU B, will only become even smaller. From this point of
view, the Russell average model seems to be a better one.
4.5 Complementary Literature 85
10
F sR– = sSBM
–
G
8
D
6
R s R–
4 C +
K sSBM s R+
+
s SBM
2 A – E
sSBM
s R+
s R– B
O 2 4 6 8 10 12 14 16 X
Fig. 4.6 Efficiency measurement of some models
4.5 Complementary Literature
The additive model has not attracted many studies, although the measure of effi-
ciency proportions (Banker & Cooper, 1994), the weighted additive model
(Steinmann & Zweifel, 2001), Bounded Adjusted Measure (BAM) (Cooper, Pastor,
Borras, et al., 2011), and decomposition in the weighted additive model (Cooper,
Pastor, Aparicio, & Borras, 2011) have been discussed in the literature. The Russell
measure has received more attention (Russell & Schworm, 2009), and many mod-
ifications have been proposed. They include modifications to the output model
(Aparicio & Pastor, 2013), weighted measures (Ruggiero & Bretschneider, 1998;
Zhu, 1996), and linear programming solutions (Briec, 2000; Sueyoshi & Sekitani,
2007), among others (Levkoff et al., 2012; Russell, 1985).
The Russell ratio measure, or the SBM model in Tone (2001), has many appli-
cations. Regarding methodologies, models have been developed to handle negative
inputs and outputs (Emrouznejad et al., 2010; Sharp et al., 2007), undesirable
outputs (Arabi et al., 2015), imprecise data (Azizi et al., 2015), stochastic data
(Azadi & Saen, 2011), context-dependent situations (Morita et al., 2005) and to
calculate super efficiencies (Chen, 2013; Fang et al., 2013; Tone, 2002). Other
studies include the incorporation of directional distance functions (Färe et al.,
2013; Färe & Grosskopf, 2010) and variants of the model in Tone (2010) for dealing
with unreasonable scores caused by remote projection points. With regard to com-
paring radial and non-radial measures, this is discussed in De Borger and Kerstens
(1996), Avkiran et al. (2008), Mahlberg and Sahoo (2011), and Mirsalehy
et al. (2014).
References
Ali, A., & Seiford, L. M. (1990). Translation invariance in data envelopment analysis. Operations
Research Letters, 9, 403–405.
Aparicio, J., & Pastor, J. T. (2013). A well-defined efficiency measure for dealing with closest
targets in DEA. Applied Mathematics and Computation, 219, 9142–9154.
Arabi, B., Munisamy, S., & Emrouznejad, A. (2015). A new slacks-based measure of Malmquist-
Luenberger index in the presence of undesirable outputs. Omega, 51, 29–37.
Avkiran, N. K., Tone, K., & Tsutsui, M. (2008). Bridging radial and non-radial measures of
efficiency in DEA. Annals of Operations Research, 164, 127–138.
Azadi, M., & Saen, R. F. (2011). Developing an output-oriented super slacks-based measure model
with an application to third-party reverse logistics providers. Journal of Multi-criteria Decision
Analysis, 18, 267–277.
Azizi, H., Kordrostami, S., & Amirteimoori, A. (2015). Slacks-based measures of efficiency in
imprecise data envelopment analysis: An approach based on data envelopment analysis with
double frontiers. Computers & Industrial Engineering, 79, 42–51.
Banker, R. D., & Cooper, W. W. (1994). Validation and generalization of DEA and its uses. Top, 2,
249–314.
Briec, W. (2000). An extended Färe-Lovell technical efficiency measure. International Journal of
Production Economics, 65, 191–199.
Charnes, A., & Cooper, W. W. (1962). Programming with linear fractional functionals. Naval
Research Logistics Quarterly, 9, 181–186.
Charnes, A., Cooper, W. W., Golany, B., Seiford, L. M., & Stutz, J. (1985). Foundations of data
envelopment analysis for Pareto-Koopmans efficient empirical production function. Journal of
Econometrics, 30, 91–107.
Chen, C. M. (2013). Super efficiencies or super inefficiencies? Insights from a joint computation
model for slacks-based measures in DEA. European Journal of Operational Research, 226,
258–267.
Cooper, W. W., Park, K. S., & Pastor, J. T. (1999). RAM: A range adjusted measure of inefficiency
for use with additive models, and relations to other models and measures in DEA. Journal of
Cooper, W. W., Pastor, J. T., Aparicio, J., & Borras, F. (2011). Decomposing profit inefficiency in
DEA through the weighted additive model. European Journal of Operational Research, 212,
411–416.
Cooper, W. W., Pastor, J. T., Borras, F., Aparicio, J., & Pastor, D. (2011). BAM: A bounded
adjusted measure of efficiency for use with bounded additive models. Journal of Productivity
Analysis, 35, 85–94.
De Borger, B., & Kerstens, K. (1996). Radial and nonradial measures of technical efficiency: An
empirical illustration for Belgian local governments using an FDH reference technology.
Emrouznejad, A., Anouze, A. L., & Thanassoulis, E. (2010). A semi-oriented radial measure for
measuring the efficiency of decision making units with negative data, using DEA. European
Fang, H. H., Lee, H. S., Hwang, S. N., & Chung, C. C. (2013). A slacks-based measure of super-
efficiency in data envelopment analysis: An alternative approach. Omega, 41, 731–734.
Färe, R., & Grosskopf, S. (2010). Directional distance functions and slacks-based measures of
efficiency. European Journal of Operational Research, 200, 320–322.
Färe, R., Grosskopf, S., & Lovell, C. A. K. (1985). The measurement of efficiency of production.
Kluwer-Nijhoff.
Färe, R., Grosskopf, S., & Whittaker, G. (2013). Directional output distance functions: Endogenous
directions based on exogenous normalization constraints. Journal of Productivity Analysis, 40,
267–269.
References 87
Färe, R., & Lovell, C. A. K. (1978). Measuring the technical efficiency of production. Journal of
Färe, R., Lovell, C. A. K., & Zieschang, K. (1983). Measuring the technical efficiency of multiple
outputs technologies. In W. Eichhorn, R. Henn, K. Neumann, & R. W. Shephard (Eds.),
Quantitative studies on production and prices. Physica-Verlag.
Levkoff, S. B., Russell, R. R., & Schworm, W. (2012). Boundary problems with the “Russell”
graph measure of technical efficiency: A refinement. Journal of Productivity Analysis, 37,
239–248.
Mahlberg, B., & Sahoo, B. K. (2011). Radial and non-radial decompositions of Luenberger
productivity indicator with an illustrative application. International Journal of Production
Economics, 131, 721–726.
Mirsalehy, A., Abu Bakar, M. R., Jahanshahloo, G. R., Lotfi, F. H., & Lee, L. S. (2014). Centralized
resource allocation for connecting radial and nonradial models. Journal of Applied Mathemat-
ics, 2014, 1–12. https://doi.org/10.1155/2014/974075
Morita, H., Hirokawa, K., & Zhu, J. (2005). A slacks-based measure of efficiency in context-
dependent data envelopment analysis. Omega, 33, 357–362.
Pastor, J. T., Ruiz, J. L., & Sirvent, I. (1999). An enhanced DEA Russell graph efficiency measure.
Ruggiero, J., & Bretschneider, S. (1998). The weighted Russell measure of technical efficiency.
Russell, R. R. (1985). Measures of technical efficiency. Journal of Economic Theory, 35, 109–126.
Russell, R. R., & Schworm, W. (2009). Axiomatic foundations of efficiency measurement on data-
generated technologies. Journal of Productivity Analysis, 31, 77–86.
Sharp, J. A., Meng, W., & Liu, W. (2007). A modified slacks-based measure model for data
envelopment analysis with ‘natural’ negative outputs and inputs. Journal of the Operational
Research Society, 56, 1672–1677.
Steinmann, L., & Zweifel, P. (2001). The range adjusted measure (RAM) in DEA: Comment.
Sueyoshi, T., & Sekitani, K. (2007). Computational strategy for Russell measure in DEA: Second-
order cone programming. European Journal of Operational Research, 180, 459–471.
Tone, K. (2001). A slacks-based measure of efficiency in data envelopment analysis. European
Tone, K. (2002). A slacks-based measure of super-efficiency in data envelopment analysis.
Tone, K. (2010). Variations on the theme of slacks-based measure of efficiency in DEA. European
Zhu, J. (1996). Data envelopment analysis with preference structure. Journal of the Operational
Chapter 5
Efficiency Measurement in Special
Production Stages
The preceding three chapters discuss how to measure the efficiency of a DMU based
on the production frontier constructed from the peer DMUs. The production possi-
bility set is assumed to be convex. In the classical production theory, production is
separated into three stages, as depicted in Fig. 5.1 (Ferguson & Gould, 1986). At the
beginning, the output rises at an increasing rate as the input increases, then at a
decreasing rate, and finally at a negative rate. The first stage corresponds to use of the
variable input X to point b, where the average product (AP) achieves its maximum.
At this point the marginal product (MP) equals the average product. Stage II starts
from this point to point c, where the marginal product of X drops to zero and the total
product (TP) of X culminates. Stage III corresponds to use of the variable input X to
the right of this point, where the marginal product is negative.
The postulate of convexity of the BCC model permits increasing, constant, and
decreasing returns to scale. At the same time, it also requires the marginal product to
be non-increasing, which excludes the region before point a in Fig. 5.1.
In the region before point a, the fixed input is being utilized more effectively as
the variable input X increases. The marginal product thus increases accordingly
(Henderson & Quandt, 1980). In contrast, the fixed input is not large enough to
accommodate the variable input X, as the latter keeps increasing after point c; the
marginal product thus becomes negative. Although the production possibility set
after point c is still convex, the conventional DEA models are not able to properly
measure the efficiency for the DMUs located in this region. The variable input is
congested when it is much larger than the fixed input, and the effect that the output
declines due to too much of the variable input is called congestion. If this effect is
ignored, then the real efficiency of the DMUs in this region will be understated. A
result of taking the phenomenon of increasing rate in marginal product into consid-
eration is the non-convexity of the production possibility set. Without taking this
effect into account, the efficiency of the DMUs in this region will also be
understated.

https://doi.org/10.1007/978-3-031-27593-7_5
90 5 Efficiency Measurement in Special Production Stages
Fig. 5.1 Three stages of

production
Stage I Stage II Stage III
TP
O a b c X
AP
X
MP
This chapter discusses these two effects and introduces some models to alleviate
their effects on efficiency measurement. They include the multiplicative and free
disposal hull models for the case of non-convex production possibility sets, and the
weak disposability, slack-measure, and input-fixing models for the case of
congestion.
5.1 Multiplicative Model
In Stage I of production the marginal product increases to the maximum and then
decreases to the point where the average product culminates. In this region the
production possibility set is not convex, and the corresponding production frontier
is difficult to construct. This section introduces the multiplicative model for
constructing a non-convex production possibility set to better measure the efficiency
of the DMUs in this region. The discussion is separated into subsections on variable
returns to scale (VRS) and constant returns to scale (CRS).
5.1 Multiplicative Model 91
5.1.1 Variable Returns to Scale
The Cobb-Douglas production function is widely used to describe the production of

an entrepreneur that utilizes multiple inputs Xi, i = 1, . . ., m, to produce a single-
output Y in the form of Y = v0 m i = 1 X i . For the simplest case of m = 1, the exponent
vi
v1 with a value greater than, equal to, and less than one represents a convex curve
with an increasing rate, a straight line with a constant rate, and a concave curve with
a decreasing rate, respectively. In higher dimensions the partial derivative with
respective to Xi is increasing, constant, and decreasing when the exponent of Xi is
greater than, equal to, and less than one, respectively. The Cobb-Douglas production
function, which is log-linear, is thus able to describe the non-convex phenomenon of
the production possibility set. A straightforward idea is to replace the linear function
in the CCR and BCC models with the log-linear function.
In Chap. 3 Banker et al. (1984) postulated convexity and free disposability to
define the production possibility set under variable returns to scale. By modifying
the linear relationship to log-linear, Banker and Maindiratta (1986) defined the
following production possibility set:
n
λ
T = ðX, Y Þ j X ijj ≤ X i , i = 1, . . . , m,
j=1
n
λ
Y rjj ≥ Y r , r = 1, . . . , s,
j=1
n
λj = 1, λj ≥ 0, j = 1, . . . , n ð5:1Þ
j=1
where Xij and Yrj are the ith input and rth output, respectively, of the jth DMU. This
set is obviously not convex. However, if we let X ij = ln X ij and Y rj = ln Y rj , then the
set T defined below is convex:
n n
T= X, Y j λj X ij ≤ X i , i = 1, . . . , m, λj Y rj ≥ Y r , r = 1, . . . , s,
j=1 j=1
n
λj = 1, λj ≥ 0, j = 1, . . . , n ð5:2Þ
j=1
Because of the strictly monotonic transformation of logarithm, there exists a one-

to-one mapping of the set T onto the production possibility set T.
Analogous to the Shephard’s distance function (Shephard, 1970) defined in the
original coordinates, a modified Shephard’s output distance measure φ can be
defined in the logarithm coordinates for an input–output vector (X0, Y0) 2 T:
max: φ
n
λ
s:t: X ijj ≤ X i0 , i = 1, . . . , m
j=1
n
λ
Y rjj ≥ φY r0 , r = 1, . . . , s ð5:3Þ
j=1
n
λj = 1
j=1
λj ≥ 0, j = 1, . . . , n:
The DMU being evaluated is efficient if the optimal objective value is φ = 1 and
is inefficient if φ > 1, with the corresponding (output) efficiency of 1/φ.
In order to identify weakly efficient DMUs, slack variables si- and sþ
r , similar to
those used in the BCC model, are added to Model (5.3) to become:
m s ε
max: φ si- sþ
r
i=1 r=1
n
λ
s:t: si- X ijj = X i0 , si- ≥ 1, i = 1, . . . , m
j=1
n
λ ð5:4Þ
Y rjj =sþ
r = φY r0 , sþ
r ≥ 1, r = 1, . . . , s
j=1
n
λj = 1
j=1
λj ≥ 0, j = 1, . . . , n,
where ε is the non-Archimedean number similar to that used in the BCC model. By
taking logarithm appropriately, Model (5.4) is transformed into the following linear
program:
m s
r
i=1 r=1
n
s:t: λj X ij þ si- = X i0 , i = 1, . . . , m
j=1
n
ð5:5Þ
λj Y rj - sþ
r = φ þ Y r0 , r = 1, . . . , s
j=1
n
λj = 1
j=1
λj , si- , sþ
r ≥ 0, j = 1, . . . , n, i = 1, . . . , m, r = 1, . . . , s:
where the carets “^” are used to represent the natural logarithm. A DMU is efficient
if and only if the corresponding φ and all si- and sþ
r are equal to zero at optimality.
The efficiency for a DMU measured from this model is
1= exp φ þ ε m
s
i=1 i
-
þ s
s
r=1 r
þ
.
The equation for the piecewise log-linear frontier can be obtained from the dual of
Model (5.5) formulated as follows, provided the solution is unique:
m s
min: vi X i0 - ur Y r0 þ α
i=1 r=1
m s
s:t: vi X ij - ur Y rj þ α ≥ 0, j = 1, . . . , n
i=1 r=1
ð5:6Þ
s
ur = 1
r=1
vi , ur ≥ ε, i = 1, . . . , m, r = 1, . . . , s
α unrestricted in sign:
Converting the objective function and the first set of constraints to their original
form, one has min. v0 m s m s
i = 1 X i0 = r = 1 Y r0 and v0 i = 1 X ij = r = 1 Y rj ≥ 1, respec-
vi ur vi ur
tively, where v0 = exp (α). The objective value is the inverse of the (output)
efficiency. If the optimal solution (v, u) is uniquely determined, then there exists
a unique supporting hyperplane at the projection point of (X0, Y0). It follows that
s ur m vi
r = 1 Y r = v0 i = 1 X i is the surface bounding the set T at the projection point of
(X0, Y0). The marginal product of input i for output r, which is the partial derivative
of yr with respective to xi, can be derived as: ∂ yr =∂ xi = vi =X i = ur =Y r in
this case.
To see how the efficiency of a DMU is measured with the multiplicative model,
consider a simple case of six DMUs, A, B, C, D, E, and F, using one-input X to
produce one-output Y. Table 5.1 contains the data and Fig. 5.2 depicts the six DMUs
on the X-Y plane. By applying Model (5.6) the same unique solution of
Table 5.1 Data and efficiency measures for the multiplicative model example
Input Output BCC Multiplicative model
DMU X Y Eff. VRS eff. CRS eff. Scale eff.
A 5 2 1 1 0.3 0.3
B 10 5 0.4167 0.4819 0.375 0.7782
C 12 16 1 1 1 1
D 15 20 1 1 1 1
E 20 22 1 1 0.8476 0.8476
F 17.5 15 0.7143 0.7127 0.6429 0.9021
Fig. 5.2 Production frontier Y

constructed from the
multiplicative model 30
25
Q
P
E
20
D
15 C F
B*
B
10 B̂
5 B
A
O 5 10 15 20 25 X
α = - 3:1296, v1 = 2:3752 is obtained for DMUs A, B, and C. The corresponding

frontier is Y = 0.04373X2.3752, as shown by the curve OP in Fig. 5.2, where
0.04373= exp(α*). DMUs A and C lie on this curve and are thus efficient. DMU
B has an efficiency score of Y B =Y = 0:4819. Clearly, the production possibility set
B
defined by this curve is not convex. The marginal product of this curve is ∂y/
∂x = 0.1039X1.3752, which is clearly increasing. In other words, the production
frontier constructed from Model (5.6) is able to describe the situation of increasing
marginal product.
By applying Model (5.6) to measure the efficiency of DMU D, the straight line
OD, with the equation of Y = 4X/3, is obtained. Similarly, when Model (5.6) is
applied to DMUs E and F, the same frontier OQ, with the equation of
Y = 8.1543X0.3313, is obtained. DMUs D and E lie on their respective frontiers
and are efficient, whereas DMU F is inefficient, with an efficiency of 0.7127. In sum,
the production frontier of this example is composed of three piecewise log-linear

curves: Y = 0.04373X2.3752 in the range of O to C, Y = 4X/3 in the range of C to D,
and Y = 8.1543X0.3313 in the range of D to E, and thereafter, as shown in Fig. 5.2.
The marginal product is increasing in the first segment, remaining constant in the
second segment, and decreasing in the third segment, in accordance with the
production theory.
If the conventional BCC model is applied, then the production frontier will be the
piecewise line segments ACDE, connecting DMUs A, C, D, and E, where the
marginal product keeps decreasing as X increases. In this case the efficiency of
DMU B, which is located in the region of increasing marginal product, will be
underestimated. Specifically, the efficiency measured from the BCC model is 0.4167
= Y B =Y B , while that measured from the multiplicative model is 0.4819
= Y B =Y . However, it should be noted that the multiplicative model may also
B
underestimate efficiency. For example, the curve OQ lies above the line segment DE
in the region where DMU F is located, which yields an efficiency of 0.7127 from the
multiplicative model and an efficiency of 0.7143 from the BCC model. Columns
four and five of Table 5.1 show the efficiencies for the six DMUs measured from the
BCC and multiplicative models, respectively.
5.1.2 Constant Returns to Scale
The discussion of efficiency measurement in the preceding subsection is under

variable returns to scale. The multiplicative model can also be constructed under
the assumption of constant returns to scale.
Recall that under constant returns to scale the production possibility set must
satisfy the postulate of ray unboundedness, which states that (X, Y)2T implies
(ωX, ωY) 2 T for ω ≥ 0, as discussed in Chap. 3. The production possibility set
TCRS, with this postulate considered, then becomes:
n n n
T CRS = ðX,YÞ j ðωX ij Þλj ≤X i , i=1, ...,m, ðωY rj Þλj ≥Y r , r =1, ...,s λj =1; ω,λj ≥0, j=1, ...,n
j=1 j=1 j=1
n
λ
= ðX,YÞ j ω X ijj ≤X i , i=1, ...,m,
j=1
n n
λ
ω Y rjj ≥Y r , r =1, ...,s, λj =1; ω,λj ≥0, j=1, ...,n
j=1 j=1
ð5:7Þ
The corresponding output-oriented model is formulated as:

max: ρ
n
λ
s:t: ω X ijj ≤ X i0 , i = 1, . . . , m
j=1
n
λ
ω Y rjj ≥ ρY r0 , r = 1, . . . , s ð5:8Þ
j=1
n
λj = 1
j=1
ω, λj ≥ 0, j = 1, . . . , n:
Adding slack variables and taking logarithm appropriately, the following linear
model is obtained:
m s
max: ρþε si- þ sþ
r
i=1 r=1
n
s:t: ωþ λj X ij þ si- = X i0 , i = 1, . . . , m
j=1
n
ωþ λj Y rj - sþ
r = ρ þ Y r0 , r = 1, . . . , s ð5:9Þ
j=1
n
λj = 1
j=1
λj , si- , sþ
r ≥ 0, j = 1, . . . , n, i = 1, . . . , m, r = 1, . . . , s
ω unrestricted in sign:
The objective value is the logarithm of the efficiency (in reciprocal form).
Compared with the VRS Model (5.5), a variable ω is added to the input and output
constraints.
To further investigate the characteristics of the production frontier constructed
from this model, we formulate the dual of Model (5.9) as follows:
m s
min: vi X i0 - ur Y r0 þ α
i=1 r=1
m s
s:t: vi X ij - ur Y rj þ α ≥ 0, j = 1, . . . , n
i=1 r=1
m s
vi - ur = 0 ð5:10Þ
i=1 r=1
s
ur = 1
r=1
vi , ur ≥ ε, i = 1, . . . , m, r = 1, . . . , s
α unrestricted in sign:
Compared with Model (5.6), the dual of the VRS model, a constraint
m
- sr = 1 ur = 0 appears. This constraint can be simplified to m
i = 1 vi i = 1 vi = 1
due to the constraint of sr = 1 ur = 1. Model (5.10) can thus be formulated, in the
original data, as:
m vi
v0 i = 1 X i0
min: s ur
r = 1 Y r0
v0 m vi
i = 1 X ij
s:t: s ur ≥ 1, j = 1, . . . , n
r = 1 Y rj
m
ð5:11Þ
vi = 1
i=1
s
ur = 1
r=1
vi , ur ≥ ε, i = 0, . . . , m, r = 1, . . . , s
where v0 = exp (α). For sr = 1 Y ur r fixed at a specific value k, the equation of

v0 m s m
i = 1 X i = r = 1 Y r = 1 implies that v0 i = 1 X i = k is an isoquant. It is well
vi ur vi
known in economic theory that a production function of the Cobb-Douglas form

of f ðXÞ = v0 m vi
i = 1 X i has increasing, constant, and decreasing returns to scale for
m
i = 1 vi being greater than, equal to, and less than one, respectively (Intriligator,
1971). Model (5.11) has a constraint of m i = 1 vi = 1, confirming that the returns to
scale for the constructed production frontier are constant.
This model is the log-linear counterpart of the CCR model, and the measured
efficiency is an aggregation of technique and scale. Denoting the efficiency mea-
sured from the VRS model as technical efficiency, the ratio of the efficiency
measured from the CRS model to that of the VRS model is scale efficiency.
The example in Table 5.1 has only one input and one output. The frontier
constructed under constant returns to scale is the straight line OD, which is the
same as that constructed from the conventional CCR model. The second-to-last
column of Table 5.1 shows the efficiencies of the six DMUs measured from the
multiplicative model under constant returns to scale. The ratio of the CRS efficiency
to the VRS efficiency is scale efficiency, as shown in the last column.
The efficiency measured from the multiplicative model has a property which is
not shared by its linear counterpart of the CCR and BCC models. Consider a case
where every DMU applies the same amounts of inputs X1, . . ., Xm to produce
different combinations of outputs Y1, . . ., Ys. Suppose there is a super DMU R that
produces the largest amount of every output, that is, XiR = Xij for i = 1, . . ., m, and
YrR ≥ Yrj for r = 1, . . ., s, j = 1, . . ., n. This implies that the efficiency of output r of a
DMU is E r0 = Y r0 =Y rR . In measuring the output efficiency of a DMU, the optimal
u vi
solution (u, v) obviously satisfies the condition of sr = 1 Y rRr =v0 m i = 1 X iR = 1 for
the constraint corresponding to the super DMU R, under either variable or constant
returns to scale. The output efficiency of the DMU can then be derived as:
s ur s ur s ur s ur

r = 1 Y r0 r = 1 Y r0 r = 1 Y r0 Y r0
E0 = vi
= vi
= s ur =
v0 m v0 m Y rR
i = 1 X i0 i = 1 X iR r = 1 Y rR r=1
s
u
= E r0 r ð5:12Þ
r=1
The output efficiency of a DMU is decomposed into the product of the efficien-
cies of all outputs raised to the power of their respective exponent ur. Since the sum
of the exponents is equal to one, sr = 1 ur = 1, the efficiency of the DMU is the
weighted geometric average of the efficiencies of the outputs.
By the same token, the input efficiency of the DMU is a log-linear combination of
the efficiencies of the inputs. It will be a weighted geometric average provided
m
i = 1 vi = 1, i.e. when the returns to scale are constant.
5.2 Free Disposal Hull
In constructing the production possibility set from the observations, one widely used
postulate is convexity, which assumes that the convex combinations of the obser-
vations also exist in the production possibility set. The production possibility set
constructed under this postulate is thus convex. A weakness of this property is that
the corresponding frontier is unable to describe the situation of increasing marginal
product. Moreover, whether the convex combinations of the observations can exist
in the real world is also questionable. One idea is thus to use only the observations,
based on the postulate of free disposability, to construct the production possibility
set, and the resulting set is the free disposal hull.
5.2 Free Disposal Hull 99
5.2.1 General Case
The free disposal hull (FDH) was proposed by Deprins et al. (1984). The idea is to
construct the production possibility set only under the postulate of free disposability,
which states that if (Xj, Yj) 2 T, then (X, Yj) 2 T for X ≥ Xj and (Xj, Y) 2 T for Y ≤ Yj,
where T is the production possibility set. The production possibility set constructed
from a set of n observations (Xj, Yj), j = 1, . . ., n is then:
n n
T = ðX, Y Þ j λj X ij ≤ X i , i = 1, . . . , m, λj Y rj ≥ Y r , r = 1, . . . , s,
j=1 j=1
n
λj = 1, λj 2 f0, 1g, j = 1, . . . , n ð5:13Þ
j=1
Since λj are binary variables, and their sum is equal to one, the n terms in the
summations of nj = 1 λj X ij and nj = 1 λj Y rj reduce to only one, to result in the desired
forms of Xij ≤ Xi and Yrj ≥ Yr, j = 1, . . ., n.
The FDH model for measuring efficiency is to find either the minimum contrac-
tion parameter θ for the inputs or the maximum expansion parameter φ for the
outputs in the production possibility set defined in Expression (5.13). Specifically,
the input model is:
m s
min: θ - ε si- þ sþ
r
i=1 r=1
n
j=1
n
λj Y rj - sþ
r = Y r0 , r = 1, . . . , s ð5:14Þ
j=1
n
λj = 1
j=1
si- , sþ
r ≥ 0, i = 1, . . . , m, r = 1, . . . , s
λj 2 f0, 1g, j = 1, . . . , n
where ε is the non-Archimedean number included to detect weakly efficient DMUs,

and the objective value is the input efficiency of the DMU. The output model is:
Table 5.2 Data and efficiency measures for the FDH example
Input Output Input efficiency Output efficiency CRS eff.
DMU X Y FDH (BCC) FDH (BCC) FDH
A 5 2 1 (1) 1 (1) 3/10
B 10 5 1 (13/20) 1 (5/12) 3/8
C 12 16 1 (1) 1 (1) 1
D 15 20 1 (1) 1 (1) 1
E 20 22 1 (1) 1 (1) 33/40
F 17.5 15 24/35 (23/35) 3/4 (5/7) 9/14
m s
max: φ þ ε si- þ sþ
r
i=1 r=1
n
s:t: λj X ij þ si- = X i0 , i = 1, . . . , m
j=1
n ð5:15Þ
λj Y rj - sþ
r = φY r0 , r = 1, . . . , s
j=1
n
λj = 1
j=1
si- , sþ
r ≥ 0, i = 1, . . . , m r = 1, . . . , s
where the objective value is the inverse of the output efficiency.

The production possibility set of the free disposal hull in the two-dimensional
space is the region defined by a step function. Consider the data contained in
Table 5.2, which is the same as that contained in Table 5.1. The production frontier
constructed from the input model (5.14), or the output model (5.15), is the stairs-
shaped line segments AA′BB′CC′DD′E, as shown in Fig. 5.3. Obviously, the
production possibility set is not convex. Moreover, it is the smallest set defined by
linear functions generated from the observations. As a result, the measured effi-
ciency is the largest among all linear DEA models. In this example, DMUs A, B,
C, D, and E are efficient, and DMU F is inefficient, with an input efficiency of
X F =X F = 12=17:5 and an output efficiency of Y E =Y F o = 15=20. Recall that DMU
B is inefficient in the multiplicative model. Columns four and six of Table 5.2 show
the input and output efficiencies of the six DMUs, respectively.
If the conventional BCC model is applied, then the frontier is the line segments
ACDE. In this case, the efficient DMU B becomes inefficient, with the input
efficiency of X =X B = 6:5=10 and output efficiency of Y B =Y = 5=12, and the
B B
efficiency of the inefficient DMU F becomes X =X F = 11:5=17:5 from the input
F
side and Y F =Y = 15=21 from the output side. The BCC efficiencies of the six
F
DMUs are shown in columns five and seven in parentheses. Clearly, the FDH model
produces higher efficiency scores.
Fig. 5.3 Production frontier Y

constructed from the FDH
model in the X-Y plane
25
~ E
F
20
D F0 D
C C
15
F̂ F F
~
B
10
5 B
B* B̂ B
A A
O 5 10 15 20 X
To better grasp the idea of the production possibility set associated with the free
disposal hull, consider a two-input one-output case with the DMUs plotted in
Fig. 5.4, where there are five DMUs applying different combinations of inputs X1
and X2 to produce one unit of output Y. The isoquant for Y = 1 is the stairs-shaped
graph of AA′BB′CD, and the production possibility set is the region to the northeast
of the isoquant, which is obviously not convex. The isoquant indicates that DMUs
A, B, and C are strongly efficient, DMU D is weakly efficient, and DMU E is
inefficient, with an efficiency score of OE=OE. If the conventional BCC model is
applied, then the isoquant will be the line segments ACD. In this case the efficient
DMU B becomes inefficient, with an efficiency measure of OB=OB, and the effi-
ciency of DMU E drops to a smaller value of OE=OE. The FDH model still produces
higher efficiency scores.
The efficiency measurement discussed in the preceding subsection is for general

cases, and this does not make any assumption regarding returns to scale. A special
case is to require the returns to scale to be constant (Kerstens & Vanden Eeckaut,
1999).
Fig. 5.4 Production frontier X2

constructed from the FDH
model in the input space
4 A A
B B
3
B̂
2 E
E
1 Ê
C D
O 1 2 3 4 5 X1
To construct the production possibility set under constant returns to scale, the
postulate of ray unboundedness, which states (ωXj, ωYj) 2 T if (Xj, Yj) 2 T, for ω ≥ 0,
is added. The inclusion of this condition changes the production possibility set to:
n n
T CRS = fðX,YÞj λj ðωX ij Þ≤X i , i= 1, ...,m, λj ðωY rj Þ≥ Y r , r = 1, ...,s,
j =1 j=1
n
λj = 1; ω≥0,λj 2 f0,1g, j= 1, ...,ng
j=1
ð5:16Þ
The associated FDH model for measuring the efficiency of a DMU from the input
side then is:
m s
r
i=1 r=1
n
s:t: λj ωX ij þ si- = θX i0 , i = 1, . . . , m
j=1
n
λj ωY rj - sþ
r = Y r0 , r = 1, . . . , s ð5:17Þ
j=1
n
λj = 1
j=1
ω, si- , sþ
r ≥ 0, i = 1, . . . , m, r = 1, . . . , s
λj 2 f0, 1g, j = 1, . . . , n
For cases of only one input and one output, the algebraic meaning of this model is
to find an ω for every Yj, such that ωYj = Y0. Then for these Yj whose corresponding
Xj multiplied by ω is less than or equal to X0, we calculate θj = ωXj/X0, and the
smallest θj is the efficiency θ.
Geometrically, the production frontier is the ray emanating from the origin
pointing at the DMU that superimposes upon all DMUs. In the one-input one-output
case, this is the same as that obtained from the CCR model. Using the DMUs in
Fig. 5.3 to explain this, the production frontier is ray OC. The efficiency of a DMU is
the ratio of the input value of its projection point on the frontier to that of itself. For
example, the (input) efficiency of DMU B is X B =X B . The last column of Table 5.2
shows the efficiencies of the six DMUs measured from Model (5.17). Similar to the
case of the conventional radial measures that the ratio of the CCR efficiency to the
BCC efficiency represents scale efficiency, the ratio of the efficiency calculated from
Model (5.17) to that calculated from Model (5.14) is the scale efficiency under the
FDH technology. Another similarity is that the input and output efficiencies mea-
sured from the FDH technology are the same.
One difficulty with Model (5.17) is the nonlinearity of the constraints, due to the
nonlinear terms λjω. Different forms of mixed integer linear programming and linear
programming models have been developed to measure the FDH efficiencies. The
following is a linear programming transformation proposed by Leleu (2006) to
measure the FDH efficiency under constant returns to scale:
n
min: θj
j=1
s:t: λj þ ωj X ij ≤ θj X i0 , i = 1, . . . , m, j = 1, . . . , n
λj þ ωj Y rj ≥ λj Y r0 , r = 1, . . . , s, j = 1, . . . , n
n
ð5:18Þ
λj = 1
j=1
λj ≥ 0, j = 1, . . . , n
ωj unrestricted in sign, j = 1, . . . , n
Note that the non-Archimedean number ε has been ignored in this model to make
the expression simpler. Each DMU in this model generates m+s constraints from the
m inputs and s outputs to result in a total of n(m+s) constraints. At optimality, there
will be only one λj which is nonzero, with a value of one. The DMU corresponding
to the nonzero λj lies on the frontier, and the θj corresponding to the nonzero λj is the
efficiency.
5.3 Congestion
Stage III of production shown in Fig. 5.1 is a very special situation, where excessive
amounts of the inputs cause a reduction in the outputs. This phenomenon is referred
to as congestion. One typical example of congestion is in mining, when too many
workers are crowded in a narrow underground pit, thus reducing the amount of ore
that can be excavated.
To get some idea of the production frontier when congestion occurs, consider five
DMUs, labeled as A, B, C, D, and E, in Fig. 5.5. The production frontier constructed
by the BCC model is the line segments connecting DMUs A, B, and C, and extending
horizontally to the right from C. In this case DMUs D and E are inefficient. However,
if the marginal product is allowed to be negative, then the production frontier that
best describes the production of the five DMUs is the line segments ABCD, which
bends downward after DMU C. The downward bending of the production frontier is
due to the effect of congestion. The extent of a DMU lying below CD is purely due to
the use of an inefficient technique, and the difference between CD and CD is due to
congestion. Y E =Y E can thus be defined as pure technical efficiency, and Y E =Y can
E
be defined as the congestion effect. From the input side, any DMU located in the
region between C and D can reduce its input to the amount of XC, while maintaining
its level of output. The difference between the input level of this DMU and that of
DMU C is the amount due to congestion. For example, the input amount of
congestion for DMU E is sc, as shown in Fig. 5.5.
5.3 Congestion 105
Fig. 5.5 Congestion effect

in the X-Y plane
Several models have been proposed to measure the effect of congestion, and three
approaches, weak disposability, slack-measure, and input-fixing, are introduced in
this section.
5.3.1 Weak Disposability Model
One postulate used to develop the BCC model is free disposability, which, from the
input side, states that “if any input is increased, then outputs do not decrease.”
Compared to other types of disposability, this is the strongest condition and is
usually referred to as strong disposability. The BCC model can be formulated
from either the input or output side. Since Cooper et al. (2000) found that the
input model may produce erroneous results in measuring the effect of congestion,
the output model is adopted for discussion in this section. To maintain continuity, the
BCC output model is reformulated here as:
min: φ
n
s:t: λj X ij ≤ X i0 , i = 1, . . . , m
j=1
n
λj Y rj ≥ Y r0 =φ, r = 1, . . . , s
j=1 ð5:19Þ
n
λj = 1
j=1
λj ≥ 0, j = 1, . . . , n
ωj unrestricted in sign, j = 1, . . . , n,
where the non-Archimedean number ε has been ignored for simplicity of expression.
Note that the conventional form of maximizing the expansion factor φ has been
formulated as minimizing its inverse, so that the objective value is the efficiency
score, rather than the reciprocal, for easy comparison.
In contrast to the BCC model, constructed under the assumption of strong
disposability of inputs, Byrnes et al. (1984) proposed the following model under
weak disposability, which states that “if all inputs are increased proportionally, then
outputs do not decrease,” to measure the effect of congestion:
min: β
n
s:t: λj X ij = τX i0 , i = 1, . . . , m
j=1
n
λj Y rj ≥ Y r0 =β, r = 1, . . . , s
j=1 ð5:20Þ
n
λj = 1
j=1
τ≤1
λj ≥ 0, j = 1, . . . , n
The decision variable τ in the first set of constraints is used to proportionally scale
the inputs. The objective value β is termed pure technical efficiency. The input
constraints of the BCC model (5.19), nj = 1 λj X ij ≤ X i0 , i = 1, . . ., m, are equivalent to
n
j = 1 λj X ij = τi X i0 for τi ≤ 1. Since Model (5.20) requires all τi to be the same, which
is more stringent, its objective value β is greater than or equal to the BCC efficiency
φ, defined in Model (5.19). The difference in technologies under weak and strong
disposability of inputs, expressed as γ = φ/β, is the measure of congestion. This
expression also indicates that the BCC efficiency φ is the product of pure technical
efficiency β and congestion effect γ. Together with the property that the CCR
efficiency is the product of the BCC efficiency and scale efficiency, as demonstrated
5.3 Congestion 107
in Chap. 2, the CCR efficiency is decomposed as: (CCR efficiency) = (pure

technical efficiency) × (congestion effect) × (scale efficiency).
One weakness of Model (5.20) is that when there is only one input, the input
constraint nj = 1 λj X j = τX 0 is the same as that of the BCC model nj = 1 λj X j ≤ X 0 . In
this case β of Model (5.20) will be the same as the BCC efficiency φ, indicating that
the congestion effect cannot be identified.
5.3.2 Slack-Measure Model
Another approach for measuring the effect of congestion, proposed by Cooper et al.
(1996), is to separate the input slack into those due to weak efficiency and congestion
in an additive form. The idea is to first find the projection point of the DMU being
evaluated by the BCC model (5.19). Let λj be the optimal solution, then the
projection point is:
n
^ i0
X = λj X ij , i = 1, . . . , m
j=1
n
Y^ r0 = λj Y rj , r = 1, . . . , s
j=1
The input slack si- = X i0 - X i0 is the aggregate effect of weak inefficiency and
congestion. By fixing the outputs at the values of the projection point, the largest
total amount of the input slacks shows the quantity between weak and strong
efficiencies:
m
max: δi-
i=1
n
s:t: λj X ij - δi- = X i0 , i = 1, . . . , m
j=1
n
λj Y rj = Y r0 , r = 1, . . . , s ð5:21Þ
j=1
n
λj = 1
j=1
δi ≤ si- , i = 1, . . . , m
δi , λj ≥ 0, i = 1, . . . , m, j = 1, . . . , n:
The aggregate slack si- minus the slack associated with weak efficiency δi- is
the slack associated with congestion, sci = si- - δi- . For example, DMU E in
Fig. 5.5 has a value of zero for δ-, and a value of CE for sc. Since there is only one
input in this example, the subscript i is not needed.
The congestion effect in this model is expressed in absolute quantity, showing the
amounts that the inputs can be reduced, while the current levels of outputs do not
decrease. Different from the weak disposability model, where the BCC efficiency is
decomposed into the product of pure technical efficiency and congestion effect, the
slack-measure model decomposes the BCC efficiency (in slacks) into the sum of
weak efficiency (in slacks) and congestion (in slacks).
Model (5.21) treats all inputs as equally important, as reflected by attaching the
same weight of one to every δi- in the objective function m -
i = 1 δi . If an input is
considered more important, then a relatively larger weight can be assigned to the
corresponding δi- . This is necessary, because an input can be measured by using
different units, for example, centimeters versus kilometers. Different units will cause
an input to be over- or under-emphasized compared to others and produce a
misleading result. Another weakness of this approach is that the congestion effect
is obtained in two steps, which is more tedious than the one-step solution of the weak
disposability approach and the input-fixing approach that is discussed in the follow-
ing subsection.
5.3.3 Input-Fixing Model
The slack-measure approach fixes the outputs at the values of the projection point to
measure the effect due to congestion. A similar idea is to fix the inputs at the values
of the projection point to measure the congestion effect.
Wei and Yan (2004) and Tone and Sahoo (2004) proposed the following model to
measure the congestion effect:
min: η
n
s:t: λj X ij = X i0 , i = 1, . . . , m
j=1
n
λj Y rj ≥ Y r0 =η, r = 1, . . . , s ð5:22Þ
j=1
n
λj = 1
j=1
λj ≥ 0, j = 1, . . . , n:
The first set of constraints is to represent the inputs of the DMU being evaluated
as a convex combination of the reference DMUs. The second set of constraints
expresses the outputs of this DMU, multiplied by an expansion factor 1/η, as the
same convex combination of the reference DMUs. The objective function is to
5.3 Congestion 109
determine the set of reference DMUs that yields the largest expansion factor 1/η
(or the smallest expansion factor η) for the output.
Model (5.22) is very similar to the BCC model (5.19), in that only the first set of
inequality constraints of “≤Xi0” is replaced with the equality constraints of “=Xi0”.
This model has a dual of the following form:
s
max: ur Y r0
r=1
m
s:t: vi X i0 þ v0 = 1
i=1
ð5:23Þ
s m
ur Y rj - vi X ij þ v0 ≤ 0, j = 1, . . . , n
r=1 i=1
ur ≥ 0, r = 1, . . . , s
vi unrestricted in sign, i = 0, . . . , m:
This model is also similar to the dual of the BCC model, except that the
multipliers vi are allowed to be negative. In the two-dimensional space this implies
that the frontier is allowed to bend downward. The line segment CD in Fig. 5.5, with
a negative slope, is an example of this.
Similar to the weak disposability model, the objective value of Model (5.22) is the
pure technical efficiency, because the effect of congestion has been taken into
account. Model (5.22) is more stringent than the BCC model (5.19), due to the
equality form of the first set of constraints. The objective value of the former is thus
greater than or equal to that of the latter. The congestion effect, which is the ratio of
the BCC efficiency φ to the pure technical efficiency η, is less than or equal to one.
Using DMU E in Fig. 5.5 to explain this, the ratio X E =X E is its pure technical
efficiency, and the ratio X E =X E is the congestion effect. Therefore, this model also
shares the merit of decomposition, in that (CCR efficiency) = (pure technical
efficiency) × (congestion effect) × (scale efficiency).
5.3.4 Comparison
Since the weak disposability model is unable to identify the congestion effect when
there is only one input, here an example of two inputs and one output is used for
comparing the three approaches in measuring the congestion effect.
Columns two to four of Table 5.3 show the hypothetical data for seven DMUs
taken from Kao (2010), and Fig. 5.6 depicts the isoquants constructed from them.
DMUs B, C, D, and E are clearly efficient (based on the BCC model), because they
lie on the isoquants. In Fig. 5.6, S1BCS1′ and S2DES2′ are the isoquants for Y = 1 and
Y = 2, respectively. As the DMUs move toward northeast of the isoquant S2DES2′,
Table 5.3 Congestion effects measured from different models

Input Output BCC Weak disposability Slack-measure Input-fixing
DMU X1 X2 Y φ β φ/β (sc1 , sc2 ) η φ/η
A 3 7 1 2/3 1 2/3 (0, 1) 1 2/3
B 2 4 1 1 1 1 (0, 0) 1 1
C 5 2 1 1 1 1 (0, 0) 1 1
D 4 8 2 1 1 1 (0, 0) 1 1
E 5 4 2 1 1 1 (0, 0) 1 1
F 7 7 1 1/2 1/2 1 (2, 3) 16/21 21/32
G 8 8 1 1/2 1/2 1 (3, 4) 1 1/2
Fig. 5.6 Isoquants and X2

congestion effects of the
two-input one-output
example S2
T
8 S1 D G (Y=1)
A
F (Y=1)
6 A
4 B S2 (Y=2)
E
T
2 S1 (Y=1)
C
O 2 4 6 8 X1
the outputs are expected to increase. However, the outputs of DMUs F and G have an
output less than two, indicating that congestion has occurred, and the production
frontier starts bending downward.
The region bounded by the isoquant S1BCS1′ exhibits strong disposability of
inputs (for Y = 1), and the region bounded by ABCT′ exhibits weak disposability of
inputs. DMUs A, F, and G are located in the region defined under strong dispos-
ability of inputs and are thus inefficient based on the BCC model. Of these three
DMUs, F and G are also located in the region defined under weak disposability, they
are thus pure technically inefficient, as viewed from the weak disposability model. In
contrast, DMU A is located on the boundary of the region defined under weak
disposability and is therefore pure technically efficient according to Model (5.20).
The pure technical efficiencies of the seven DMUs measured from Model (5.20) are
shown in column six of Table 5.3. Compared with the BCC efficiencies shown in
column five, only DMU A is found to have any congestion effect. However, it is
clear that DMUs F and G also have congestion effects, only that they have not been
detected. Column seven shows the congestion effect measured from this model.
When the slack-measure model is applied, DMUs A, F, and G are found to have
the congestion effect, with the slacks associated with congestion shown in column
eight. The sci value indicates the amount of input i that can be reduced while still
maintaining the current levels of outputs. For the three congested DMUs, A, F, and
G, they will move to points A′, E, and E, respectively, when the corresponding sci
values are reduced.
The input-fixing model (5.22) shows that all DMUs, except F, are pure techni-
cally efficient. Different from the slack-measure model, which shows the amounts of
inputs that can be reduced to maintain the current levels of outputs, the input-fixing
model shows the expected amounts of outputs that can be produced when congestion
occurs. In this example, the output of DMU F is expected to increase to
1/(16/21) = 21/16 if it becomes pure technically efficient. For the three congested
DMUs, A, F, and G, their outputs are expected to expand by 3/2, 32/21, and 2/1
times, respectively, if they are not congested. The last two columns show the pure
technical efficiency and congestion effects of the seven DMUs measured from this
model.
In addition to the three aforementioned approaches, other models have also been
proposed, as reviewed in Khodabakhshi et al. (2014).
The study of multiplicative models can be traced back to the work of Banker et al.
(1981), although in this pioneering work the outputs do not compete for the inputs.
Charnes et al. (1982, 1983) elaborated on this, and Banker and Maindiratta (1986)
then provided a theoretical foundation for the approach. The log-linear technology
was then applied to develop a multiplicative directional distance function to measure
technical efficiency by Mehdiloozad et al. (2014).
The free disposal hull models are conventionally represented in the mixed integer
programming form (Podinovski, 2004a). They can be solved via enumeration
algorithms (Briec et al., 2004; Cherchye et al., 2001a; Kerstens & Van De Woestyne,
2014; Keshvari & Hardoroudi, 2008; Tulkens, 1993) or reformulated as linear
programming models for solutions (Agrell & Tind, 2001; Kerstens & Vanden
Eeckaut, 1999; Leleu, 2006). Another stream of studies concentrates on returns to
scale, such as Briec et al. (2000), Podinovski (2004b), Soleimani-damaneh et al.
(2006), Soleimani-damaneh and Mostafaee (2009), De Witte and Marques (2011).
The free disposal hull models have also been modified to construct different
non-convex production possibility sets to measure efficiencies (Blancard et al.,

2011; Bogetoft, 1996; Bogetoft et al., 2000; Leleu, 2009; Petersen, 1990).
The discussion of congestion started with the work of Färe and Svensson (1980).
Later, Färe and Grosskopf (1983) and Färe et al. (1985) developed a data envelop-
ment analysis model to measure the congestion effect, while Byrnes et al. (1984) and
Byrnes et al. (1988) decomposed the BCC efficiency into the product of pure
technical efficiency and congestion effect, based on the assumption of weak dispos-
ability of inputs. McDonald (1996) and Färe and Grosskopf (2000) investigated the
problems in the decomposition, while Cherchye et al. (2001b), Färe and Grosskopf
(2001), and Cooper et al. (2002) discussed the characteristics of different
approaches. Sharma and Yu (2013) developed a multi-stage DEA model to measure
the efficiency and congestion of a supply chain. Zare-Haghighi et al. (2014) devel-
oped a non-radial efficiency measure to incorporate undesirable outputs, and
Khodabakhshi et al. (2014) presented a review of the methods for estimating input
congestions in DEA.
References
Agrell, P. J., & Tind, J. (2001). A dual approach to nonconvex frontier models. Journal of
Productivity Analysis, 16(2), 129–147.
Banker, R. D., Charnes, A., Cooper, W. W., & Schinnar, A. (1981). A bi-extremal principle for
frontier estimation and efficiency evaluation. Management Science, 27, 1370–1382.
Banker, R. D., & Maindiratta, A. (1986). Piecewise loglinear estimation of efficient production
surfaces. Management Science, 32, 126–135.
Blancard, S., Boussemart, J. P., & Leleu, H. (2011). Measuring potential gains from specialization
under non-convex technologies. Journal of the Operational Research Society, 62, 1871–1880.
Bogetoft, P. (1996). DEA on relaxed convexity assumptions. Management Science, 42, 457–465.
Bogetoft, P., Tama, J. M., & Tind, J. (2000). Convex input and output projections of nonconvex
production possibility sets. Management Science, 46, 858–869.
Briec, W., Kerstens, K., Leleu, H., & Vanden Eeckaut, P. (2000). Returns to scale information on
nonparametric deterministic technologies: A simplification of goodness-of-fit methods. Journal
of Productivity Analysis, 14(3), 267–274.
Briec, W., Kerstens, K., & Vanden Eeckaut, P. (2004). Non-convex technologies and cost func-
tions: Definitions, duality and nonparametric tests of convexity. Journal of Econometrics, 81(2),
155–192.
Byrnes, P., Färe, R., & Grosskopf, S. (1984). Measuring productive efficiency: An application to
Illinois strip mines. Management Science, 30, 671–681.
Byrnes, P., Färe, R., Grosskopf, S., & Lovell, C. A. K. (1988). The effect of union on productivity:
US surface mining of coal. Management Science, 34, 1037–1053.
Charnes, A., Cooper, W. W., Seiford, L., & Stutz, J. (1982). A multiplicative model for efficiency
analysis. Socio-Economic Planning Sciences, 16, 223–224.
Charnes, A., Cooper, W. W., Seiford, L., & Stutz, J. (1983). Invariant multiplicative efficiency and
piecewise Cobb-Douglas envelopments. Operations Research Letter, 2(3), 1010–1013.
Cherchye, L., Kuosmanen, T., & Post, T. (2001a). FDH directional distance functions: With an
application to European commercial banks. Journal of Productivity Analysis, 15, 201–215.
References 113
Cherchye, L., Kuosmanen, T., & Post, T. (2001b). Alternative treatments of congestion in DEA: A
rejoinder to Cooper, Gu, and Li. European Journal of Operational Research, 132, 75–80.
Cooper, W. W., Deng, H., Huang, Z. M., & Li, S. X. (2002). A one-model approach to congestion
in data envelopment analysis. Socio-Economic Planning Sciences, 36, 231–238.
Cooper, W. W., Seiford, L. M., & Zhu, J. (2000). A unified additive model approach for evaluating
inefficiency and congestion with associated measures in DEA. Socio-Economic Planning
Sciences, 34, 1–25.
Cooper, W. W., Thompson, R. G., & Thrall, R. M. (1996). Introduction: Extension and new
developments in DEA. Annals of Operations Research, 66, 3–45.
Deprins, D., Simar, L., & Tulkens, H. (1984). Measuring labor efficiency in post offices. In
M. Marchand, P. Pestieu, & H. Tulkens (Eds.), The performance of public enterprises: Concepts
and measurements. North Holland.
De Witte, K., & Marques, R. C. (2011). Big and beautiful? On non-parametrically measuring scale
economies in non-convex technologies. Journal of Productivity Analysis, 35, 213–226.
Färe, R., & Grosskopf, S. (1983). Measuring congestion in production. Nationalökonomie, 43,
257–271.
Färe, R., & Grosskopf, S. (2000). Decomposing technical efficiency with care. Management
Science, 46, 167–168.
Färe, R., & Grosskopf, S. (2001). When can slacks be used to identify congestion? An answer to
W.W. Cooper, L. Seiford and J. Zhu. Socio-Economic Planning Sciences, 35, 217–221.
Färe, R., Grosskopf, S., & Lovell, C. A. K. (1985). The measurement of efficiency of production.
Kluwer-Nijhoff.
Färe, R., & Svensson, L. (1980). Congestion of production factors. Econometrica, 48, 1745–1753.
Ferguson, C. E., & Gould, J. P. (1986). Microeconomic theory (6th ed.). Irwin.
Henderson, J. M., & Quandt, R. E. (1980). Microeconomic theory (3rd ed.). McGraw-Hill.
Intriligator, M. D. (1971). Mathematical optimization and economic theory. Prentice-Hall.
Kao, C. (2010). Congestion measurement and elimination under the framework of data envelop-
ment analysis. International Journal of Production Economics, 123, 257–265.
Kerstens, K., & Vanden Eeckaut, P. (1999). Estimating returns to scale using non-parametric
deterministic technologies: A new method based on goodness-of-fit. European Journal of
Operational Research, 113, 206–214.
Kerstens, K., & Van De Woestyne, I. (2014). Solution methods for nonconvex free disposal hull
models: A review and some critical comments. Asia-Pacific Journal of Operational Research,
31. https://doi.org/10.1142/S0217595914500109
Keshvari, A., & Hardoroudi, N. D. (2008). An extended numeration method for solving free
disposal hull models in DEA. Asia-Pacific Journal of Operational Research, 25, 689–696.
Khodabakhshi, M., Hosseinzadeh, F., & Aryavash, K. (2014). Review of input congestion estimat-
ing methods in DEA. Journal of Applied Mathematics. https://doi.org/10.1155/2014/963791
Leleu, H. (2006). A linear programming framework for free disposal hull technologies and cost
functions: Primal and dual models. European Journal of Operational Research, 168, 340–344.
Leleu, H. (2009). Mixing DEA and FDH models together. Journal of the Operational Research
Society, 60, 1730–1737.
McDonald, J. (1996). A problem with the decomposition of technical inefficiency into scale and
congestion components. Management Science, 42, 473–474.
Mehdiloozad, M., Sahoo, B. K., & Roshdi, I. (2014). A generalized multiplicative directional
distance function for efficiency measurement in DEA. European Journal of Operational
Research, 232, 679–688.
Petersen, N. C. (1990). Data envelopment analysis on a relaxed set of assumptions. Management
Science, 36, 305–314.
Podinovski, V. V. (2004a). On the linearisation of reference technologies for testing returns to scale
in FDH models. European Journal of Operational Research, 152, 800–802.
Podinovski, V. V. (2004b). Local and global returns to scale in performance measurement. Journal
of the Operational Research Society, 55, 170–178.
Sharma, M. J., & Yu, S. J. (2013). Multi-stage data envelopment analysis congestion model.
Soleimani-damaneh, M., Jahanshahloo, G. R., & Reshadi, M. (2006). On the estimation of returns
to scale in FDH models. European Journal of Operational Research, 174, 1055–1059.
Soleimani-damaneh, M., & Mostafaee, A. (2009). Stability of the classification of returns to scale in
FDH models. European Journal of Operational Research, 196, 1223–1228.
Tone, K., & Sahoo, B. K. (2004). Degree of scale economies and congestion: A unified DEA
approach. European Journal of Operational Research, 158, 755–772.
Tulkens, H. (1993). On FDH efficiency: Some methodological issues and application to retail
banking, courts, and urban transit. Journal of Productivity Analysis, 4(1), 183–210.
Wei, Q. L., & Yan, H. (2004). Congestion and returns to scale in data envelopment analysis.
Zare-Haghighi, H., Rostamy-Malkhalifeh, M., & Jahanshaloo, R. G. (2014). Measurement of
congestion in the simultaneous presence of desirable and undesirable outputs. Journal of
Applied Mathematics. https://doi.org/10.1155/2014/512157
Chapter 6
Special Types of Input and Output Factors
The DEA methodology concerns the efficiency measurement of production systems

that apply multiple inputs to produce multiple outputs. Under normal conditions, it is
desirable to have less input consumed and more output produced, because this leads
to higher efficiencies. With regard to inefficient DMUs, improvement targets can be
obtained to make them more efficient. However, there are cases in which the
decision maker has no control over the input and output factors, in that the amounts
of some factors cannot be adjusted at the discretion of the decision maker. The
conventional DEA models introduced in the preceding chapters are either unable to
handle or will produce misleading results for these cases. This chapter discusses two
of the cases, non-discretionary and undesirable factors.
The non-discretionary factors refer to the factors whose values are exogenously
fixed, which cannot be adjusted by the decision maker. Both the input and output can
be non-discretionary. An example of such an input factor is the population size. In
measuring the efficiency of the branches of a bank in different cities, the deposits
collected are affected by the population of the city (or the district), with a larger
population size generally attracting more deposits. Unfortunately, the population
size is not controllable by either the manager of the branch or the top management of
the bank. An example of an output factor that is non-discretionary is the growth of a
forest. More volume of wood yielded from a forest is generally more desired.
However, the growth in yield is governed by the growth function of the species of
the tree, and the decision maker has no way to control this. How to appropriately
measure the efficiency of a DMU, especially to achieve perfect efficiency by
adjusting only the discretionary factors, will be discussed in the first section.
The undesirable factors are a special type for which, from the output side, a
greater amount produced with the same amount of inputs is less desirable, or, from
the input side, less amount consumed with the same amount of outputs being
produced is less desired. An example of an undesirable output is the pollution
resulting from a manufacturing process, where the less pollutant that is produced
the better. An example of an undesirable input is the waste used to produce energy in
https://doi.org/10.1007/978-3-031-27593-7_6
116 6 Special Types of Input and Output Factors
an incineration plant, where more waste being burnt is more desirable. This topic is
becoming more important as the environmental issues are of growing concern and
will be discussed in the second section.
6.1 Non-discretionary Factors
One merit of the DEA technique is that it is able to identify targets for inefficient
DMUs to become efficient. From the input side this is based on the reduction of the
inputs, while maintaining the current levels of outputs, and from the output side it is
the expansion of the outputs under the current levels of inputs. When some of the
inputs or outputs are uncontrollable, the conventional way of identifying a target that
requires all input or output factors to be changed to make an improvement will be
unrealistic, and this also implies that the corresponding efficiency measure is
incorrect. Banker and Morey (1986) introduced the first model that takes the
non-discretionary phenomenon into consideration in measuring efficiency and deter-
mining the target. The first two subsections discuss the models from the input and
output sides separately, the third subsection provides some interpretation from the
dual point of view, and the fourth subsection addresses the issue of constant returns
to scale.
6.1.1 Input Model
Consider a simple case of six DMUs labeled as A, B, C, D, E, and F in Fig. 6.1. These
six DMUs use different combinations of two inputs, XD and XF, to produce one unit
of output Y, where XD is the conventional discretionary input whose amount can be
adjusted, and XF is a non-discretionary input whose amount is exogenously fixed.
The piecewise line segments CDE are the isoquant, and DMUs C, D, and E, lying on
the isoquant, are strongly efficient. DMU B, lying on the vertical line extending
upward from DMU C, and DMU F, lying on the horizontal line extending eastward
from DMU E, are weakly efficient. DMU A, lying in the interior of the input set, is
inefficient, with an (input) efficiency of OA=OA, measured from the BCC model.
The projection point A, with the discretionary and non-discretionary inputs at the
levels of X D and X F , respectively, is the target for the inefficient DMU A to become
A A
efficient. However, since the quantity of the non-discretionary input XF cannot be
altered, it is impossible for DMU A to move to the location of A. This also implies
that using A as the benchmark to measure the efficiency of DMU A is inappropriate.
The target for a DMU to become efficient is not usually unique, and any point on
the frontier that dominates this DMU can be adopted as the target for both measuring
the efficiency and making improvement. For example, all the points on the line
segments ADA of the isoquant dominate DMU A, because less amounts of the two
6.1 Non-discretionary Factors 117
Fig. 6.1 Efficiency XF

measurement with
non-discretionary input XF X BF B
F
X C
C
A A
X AFˆ D Â
~
A E F
O X AD X ADˆ X AD X ED X FD
XD
inputs than DMU A have been used to produce the same amount of the output.
Theoretically, any point on this portion of the isoquant can be adopted to measure the
efficiency of DMU A. Of all the points on ADA, only A uses the same amount of the
non-discretionary input as DMU A does. It is thus a suitable benchmark to compare
with DMU A, and its efficiency is the ratio of X D A
=X D
A . This idea is also applicable to
the two weakly efficient DMUs B and F. The target for DMU F will be DMU E, with
an efficiency of X D E =X F . In this case DMU F is actually inefficient, instead of weakly
D
efficient. The target for DMU B is itself, because its exogenously fixed input cannot
be altered. However, since this DMU is dominated by DMU C, it is weakly efficient,
with the efficiency of 1 - ε × sB, where sB = X FB - X FC denotes the slack between
DMUs B and C in XF.
This idea of fixing the quantity of the non-discretionary factor at the observed
level and searching for the minimum contraction ratio of the discretionary factor to
be the efficiency can be extended to higher dimensions. Let the first d inputs be
discretionary and the last m-d inputs be non-discretionary. The input constraints of
n -
j = 1 λj X ij þ si = θX i0 , i = 1, . . ., m, in the conventional BCC input model (2.10)
are separated into two parts: nj = 1 λj X ij þ si- = θX i0 , i = 1, . . ., d, for the discre-
tionary inputs and nj = 1 λj X ij þ si- = X i0 , i = d + 1, . . ., m, for the non-discretionary
ones. The complete model is:
m s
r
i=1 r=1
n
s:t: λj X ij þ si- = θX i0 , i = 1, . . . , d
j=1
n
λj X ij þ si- = X i0 , i = d þ 1, . . . , m
j=1 ð6:1Þ
n
λj Y rj - sþ
r = Y r0 , r = 1, . . . , s
j=1
n
λj = 1
j=1
λj , si- , sþ
r ≥ 0, j = 1, . . . , n, i = 1, . . . , m, r = 1, . . . , s:
If an assumption of constant returns to scale is imposed, then the convexity

constraint of nj = 1 λj = 1 is deleted. After an optimal solution (θ, λ, s*) is obtained,
there are three cases to discuss: θ* = 1 and all slack variables are zero, θ* = 1 and at
least one slack variable is not zero, and θ* < 1. These three cases indicate that the
corresponding DMU is strongly efficient, weakly efficient, and inefficient, with
n
j = 1 λj X ij , i = 1, . . ., d, as the target values of the discretionary inputs for the
inefficient DMU to become efficient.
The difference between the conventional BCC model and the non-discretionary
factor model (6.1) is that the term θXi0 in the input constraints for i = d + 1, . . ., m is
replaced by Xi0. Since Xi0 is greater than or equal to θXi0, the feasible region defined
by Model (6.1) is larger than that defined by the BCC model. The efficiency
measured from the non-discretionary factor model is thus less than or equal to that
measured from the conventional BCC model.
6.1.2 Output Model
The model for measuring the efficiency of the DMUs with non-discretionary output
factors can be formulated similarly based on the BCC output model (2.14). Suppose
the first d outputs are discretionary and the last s–d are non-discretionary. The output
model is:
m s
r
i=1 r=1
n
s:t: λj X ij þ si- = X i0 , i = 1, . . . , m
j=1
n
λj Y rj - sþ
r = φY r0 , r = 1, . . . , d
j=1 ð6:2Þ
n
λj Y rj - sþ
r = Y r0 , r = d þ 1, . . . , s
j=1
n
λj = 1
j=1
λj , si- , sþ
r ≥ 0, j = 1, . . . , n, i = 1, . . . , m, r = 1, . . . , s:
The efficiency is the inverse of the objective function. The strongly efficient,
weakly efficient, and inefficient cases are identified using a process similar to that
used with the input model. The target values for the discretionary outputs are also
obtained similarly, as nj = 1 λj Y rj , r = 1, . . ., d.
Consider four DMUs, A, B, C, and D, applying one unit of input X to produce
different combinations of outputs YD and YF, as shown in Fig. 6.2, where the former
is discretionary and the latter is exogenously fixed. The (YD, YF) coordinates of the
four DMUs are A = (1.5, 1), B = (1, 4), C = (3, 3), and D = (4, 2). The line segments
BCD are the product transformation curve for X = 1, which indicate that DMUs B, C,
and D are efficient. Since YF is exogenously fixed, the inefficient DMU A will be
Fig. 6.2 Efficiency YF

measurement with
non-discretionary output YF
B
4
3 C
Â
2 D
1 A
A
O 1 2 3 4 YD
compared with point A, on the vertical line emanating downward from DMU D,
rather than the radial point A, in measuring efficiency. However, since point A is
dominated by DMU D, the latter will be the benchmark for DMU A to measure
efficiency. Based on Model (6.2), the efficiency of DMU A is 1/(8/3 + ε).
6.1.3 Dual Model Interpretation
Similar to the conventional BCC model that has ratio and envelopment forms when
all factors are discretionary, the non-discretionary factor model also has these two
forms, and they are the dual of each other. The dual of the envelopment form of the
output model (6.2) is:
m s
i=1 r = dþ1
d
s:t: ur Y r0 = 1
r=1 ð6:3Þ
m s
vi X ij þ v0 - ur Y rj ≥ 0, j = 1, . . . , n
i=1 r=1
vi , ur ≥ ε, i = 1, . . . , m, r = 1, . . . , s
d
The first constraint of r = 1 ur Y r0 = 1 can be combined with the objective
function as the denominator, and the objective of minimizing a fractional term can
be changed to maximizing its inverse. The dual problem then becomes:
d
r = 1 ur Y r0
max: m
v X
i = 1 i i0 þ v0 - sr = dþ1 ur Y r0
s
r = 1 ur Y rj ð6:4Þ
s:t: m ≤ 1, j = 1, . . . , n
i = 1 vi X ij þ v0
vi , ur ≥ ε, i = 1, . . . , m, r = 1, . . . , s
To give an explanation of this model, recall that the production function of

m inputs and s outputs has the implicit form of F(X, Y) = 0. An explicit form is
s m
r = 1 ur Y r = i = 1 vi X i þ v0 , where the left-hand side of the equation is an aggre-
gate output and the right-hand side is an aggregate input. For efficient DMUs the
aggregate input is equal to the aggregate output. The expected aggregate output of a
m
DMU with observation (X0, Y0) is i = 1 vi X i0 þ v0 . The ratio of the observed
aggregate output to the expected aggregate output, sr = 1 ur Y r0 = m
i = 1 vi X i0 þ v0 ,
is the (output) efficiency of this DMU. Since the values of the (s–d )
non-discretionary outputs have been fixed, they have no effects on calculating the
efficiency, and can be excluded from both sides of the equation, to obtain
d m s
r = 1 ur Y r0 = i = 1 vi X i0 þ v0 - r = dþ1 ur Y r0 . The efficiency of the DMU, when
only the discretionary outputs are considered, then becomes
d m s
u Y
r = 1 r r0 = v X
i = 1 i i0 þ v 0 - u Y
r = dþ1 r r0 . The objective function of Model
(6.4) is to find the most favorable multipliers (u, v) to calculate efficiency, and the
constraints require that no DMU will have an efficiency greater than one, when the
multipliers selected by the focal DMU are applied to calculate the efficiency of
all DMUs.
By the same token, the input-oriented ratio model can be formulated from the
dual of the input model (6.2) as:
s m
r = 1 ur Y r0 - u0 - i = dþ1 vi X i0
max: d
i = 1 vi X i0
s
r = 1 ur Y rj - u0 ð6:5Þ
s:t: m ≤ 1, j = 1, . . . , n
i = 1 vi X ij
vi , ur ≥ ε, i = 1, . . . , m, r = 1, . . . , s
The geometric interpretation is similar to the output case.
The non-discretionary factor models (6.1) and (6.2) are constructed under variable
returns to scale (VRS). They can also be constructed under constant returns to scale
(CRS), and the idea is similar. Since the CCR model is the CRS counterpart of the
BCC model, the non-discretionary factor model under CRS can be constructed by
separating the input constraints of the CCR model into two parts and attaching the
distance parameter θ only to the discretionary inputs. Based on the CCR input model
(2.4), the corresponding non-discretionary factor model is:
m s
min: θ-ε si- þ sþ
r
i=1 r=1
n
s:t: λj X ij þ si- = θX i0 , i = 1, . . . , d
j=1
n ð6:6Þ
λj X ij þ si- = X i0 , i = d þ 1, . . . , m
j=1
n
λj Y rj - sþ
r = Y r0 , r = 1, . . . , s
j=1
λj , si- sþ
r ≥ 0, j = 1, . . . , n, i = 1, . . . , m, r = 1, . . . , s:
This model is the same as Model (6.1), the one under VRS, except the convexity
constraint of nj = 1 λj = 1 is deleted.
By the same token, the output model can be constructed from Model (6.2), with
the convexity constraint of nj = 1 λj = 1 deleted. Different from the conventional
CCR model, where the input efficiency is the same as the output efficiency, they are
different in the non-discretionary model. This is because the non-discretionary
inputs have been ignored in calculating the contraction parameter θ in the input
model, and the non-discretionary outputs are ignored in calculating the expansion
parameter φ in the output model, which makes the two models asymmetric.
The ratio form of the non-discretionary factor model can be constructed from the
dual of Model (6.6) as:
s m
r = 1 ur Y r0 - i = dþ1 vi X i0
max: d
i = 1 vi X i0
s
r = 1 ur Y rj
ð6:7Þ
s:t: m ≤ 1, j = 1, . . . , n
i = 1 vi X ij
vi , ur ≥ ε, i = 1, . . . , m, r = 1, . . . , s
The geometric interpretation of this model is similar to that of Model (6.5), the
one under VRS.
The efficiency calculated from Model (6.6) is an aggregation of technique and
scale, whereas that calculated from Model (6.1) is pure technique. The ratio of the
former to the latter is then the scale efficiency.
The model of Banker and Morey (1986) projects the inefficient DMUs onto the
frontier constructed from the DMUs with the same levels of the non-discretionary
factors. The effects of the non-discretionary factors on efficiency are neglected. If the
true efficiency is correlated with the non-discretionary factors, then the efficiency
measured from the Banker-Morey model would be overstated (Ruggiero, 2004).
Section 6.3 lists several related studies.
6.2 Undesirable Factors 123
6.2 Undesirable Factors
In the process of production, a factor is considered as undesirable if, from the output
side, a greater amount being produced is less desirable, or, from the input side, a less
amount being consumed is less desirable. Typical examples of undesirable outputs
are water and air pollution, while the waste used to produce energy in incineration
plants is an undesirable input. Undesirable outputs have long been discussed in
production economics, with parametric models being developed for performance
measurement (see, for example, Pittman, 1983). However, it was until Färe et al.
(1989) that a nonparametric DEA model was proposed. In this section several
representative approaches for dealing with undesirable factors are introduced,
including input–output exchange, data transformation, weak disposability, and
slacks-based measures.
6.2.1 Input–Output Exchange Approach
The major characteristic of an undesirable factor is that the conventional directions

of increasing the outputs and decreasing the inputs have opposite effects. From this
characteristic it seems reasonable to treat an undesirable input as an output, and an
undesirable output as an input, so that they have the expected direction. For example,
the conventional principle of production states it is preferable when more output
produced and less inputs are consumed. However, when a larger amount of unde-
sirable output being less preferred violates the principle of production, it becomes
reasonable if this factor is treated as an input.
Consider four DMUs, A, B, C, and D, where each one uses an input X to produce
one desirable output YD and one undesirable output YU,with the (X; YU, YD) obser-
vations of
A = ð1; 4, 1Þ, B = ð2:5; 2, 1Þ, C = ð3; 2, 1Þ, D = ð4; 1, 1Þ ð6:8Þ
Figure 6.3 shows the coordinates of the four DMUs on the YD = 1 plane. Since a
smaller amount of YU is preferred and YU can be increased only accompanied with
the decrease of input X to be Pareto optimal, the line segments ABD constitute the
frontier. This frontier is exactly the same as the isoquant of treating YU as an input. In
this example, the DMUs A, B, and D are efficient, and DMU C is inefficient, with an
efficiency measure of OC=OC. If DMU C wishes to become efficient, then both the
input and the undesirable output must be reduced from the levels of C to those of C.
By the same token, an undesirable input can be treated as an output in measuring the
efficiency. After the undesirable outputs are changed to inputs and the undesirable
inputs are changed to outputs, the conventional DEA models are then applied to
measure efficiency.
Fig. 6.3 Treating the YU

undesirable output as an
input
4 A
2 B C
Ĉ
1
D
O 1 2 3 4 X
6.2.2 Data Transformation
Another straightforward way of measuring the efficiency of a DMU with undesirable

factors is to transform the data to have the desired direction, and then to apply the
conventional DEA models to the transformed data. There are several forms of
transformation, such as inverse transformation, negative transformation, and shifted
negative transformation, as explained below.
6.2.2.1 Inverse Transformation
Mathematically, the order of a set of numbers will be reversed if one takes the
inverse of them (Golany & Roll, 1989), i.e. 1=X U U
j and 1=Y j . The undesirable factors
will then have the desired direction after the transformation.
Using the data in Eq. (6.8) as an example, the data after the inverse transformation
of the undesirable output YU is in the form of (X; 1/YU, YD), where A = (1; 1/4, 1),
B = (2.5; 1/2, 1), C = (3; 1/2, 1), and D = (4; 1, 1). Figure 6.4 depicts the four DMUs
on the YD = 1 plane. The production frontier AD indicates that not only DMU C, but
also DMU B, the one being evaluated as efficient by the input–output exchange
approach, are inefficient, with efficiency scores of XE/XB = 4/5 and XE/XC = 2/3,
respectively. The reason that DMU B becomes inefficient is because the inverse
transformation is not linear, while the production possibility set is composed of
linear combinations (or convex combinations, to be exact) of the DMUs, which
transform a boundary point to an interior one. As noted in Scheel (2001), the set of
Fig. 6.4 Inverse 1

transformation may fail to
identify efficient DMUs YU
1
D
1 E
2 B C
XE XB XC
O 1 2 3 4 X
efficient DMUs from the inverse transformation is contained in that from the input–
output exchange approach.
6.2.2.2 Negative Transformation
Another idea of changing larger numbers to smaller ones is to take negatives,

i.e. - X U
j and - Y j (Koopmans, 1951). For Dataset (6.8), the negative transforma-
U
j , Y ) is A=(1; –4, 1), B=(2.5; –2, 1), C=(3; –2, 1), and
tion in the form of (X; - Y U D
D=(4; –1, 1). Their positions on the YD = 1 plane, as shown in Fig. 6.5, indicate that
DMUs A, B, and D are efficient. This result is the same as that of the input–output
exchange approach.
Conceptually, the negative transformation is an ideal one, because it is linear and
symmetric (to zero). However, the negative numbers make the conventional DEA
models intractable, and this means that another transformation, which changes the
negative numbers to positive ones, is necessary. Some approaches for dealing with
negative numbers will be introduced in the next chapter.
6.2.2.3 Shifted Negative Transformation
Since it is difficult to use negative numbers with conventional DEA models to

measure efficiency, one idea is to add a sufficiently large constant to make the
negative numbers positive.
When a constant is added to the negative numbers the production frontier shifts
from the negative quadrant to the positive one, without altering its shape, and the
efficient and inefficient DMUs remain the same. For example, if a constant of five is
added to - Y Uj for the DMUs in Fig. 6.5, the production frontier ABD will be shifted
Fig. 6.5 Frontier obtained 1 2 3 4 X

from the negative O
transformation
1
D
2
B C
4 A
YU
upward for five units to the first quadrant, with the efficient DMUs A, B, and D, and
inefficient DMU C remaining the same. As long as the constant being added to - Y U j
is large enough, the frontier will be translated to the first quadrant. Seiford and Zhu
(2002) suggested using the largest value of the undesirable factor from all DMUs as
the constant. After the negative numbers are changed to positive ones, the conven-
tional DEA models are applied to measure the efficiency.
This idea can be generalized to cases with more than one undesirable factor. As
noted by Ali and Seiford (1990), an efficient DMU is always efficient, regardless of
the value of the constant being added to make the negative numbers positive, and, as
proved in Scheel (2001), the efficiencies of the inefficient DMUs increase with the
value of the constant. However, one undesirable phenomenon is that different
constants may result in different orders of efficiencies for the same set of DMUs
(Zanella et al., 2015).
6.2.3 Weak Disposability Approach
The conventional efficiency measurement assumes the outputs to be disposable; that

is, they can be discarded freely. While this postulate is reasonable for desirable
outputs, it is questionable for undesirable ones, as the undesirable outputs should not
be disposed of without cost. In this regard, Färe et al. (1989) assumed that outputs are
weakly disposable and developed a model to measure the efficiency when there are
undesirable factors.
Recall that in Chap. 3 we defined the output set as P(X) = {Y j Y ≥ 0, which can
be produced by X ≥ 0}. The production possibility set defined under the
Fig. 6.6 Output sets under

strong and weak Y2
disposability (Y D)
T C
6
4 D
2 ~ Â E
A ϕ d = (–2, 1)
1ϕ
A
T’
O 2 4 6 8 Y1 (Y U)
conventional strong disposability is that if Y 2 P(X), then Y′ 2 P(X), for Y′ ≤ Y. The

corresponding output set is
n n
PS ðX Þ = Yj λj X j ≤ X, λj Y j ≥ Y, λj ≥ 0, j = 1, . . . , n
j=1 j=1
under constant returns to scale. For weak disposability, if Y 2 P(X), then τY 2 P(X),
for 0 ≤ τ ≤ 1. The corresponding output set is
n n
PW ð X Þ = Yj λj X j ≤ X, τ λj Y j = Y, 0 ≤ τ ≤ 1, λj ≥ 0, j = 1, . . . , n
j=1 j=1
By imposing weak disposability on outputs, it is assumed that the undesirable

outputs are by-products of the desirable outputs, and they may be reduced only with
the proportional reduction of the desirable outputs.
To see the difference between strong and weak disposability, refer to Fig. 6.6,
where five DMUs A, B, C, D, and E apply the same amount of input X to produce
different combinations of outputs Y1 and Y2. The line segments CDE are the product
transformation curve. The output set under the assumption of strong disposability is
the region bounded by OTCDET′O. If the outputs (Y1, Y2) are weakly disposable and
Y2 by itself is strongly disposable, then the region is OBCDET′O. Conversely, if Y1
by itself is strongly disposable, then the region is OTCDEO.
For general cases of b undesirable outputs, Yr, r = 1, . . ., b, and s–b desirable
outputs, Yr, r = b + 1, . . ., s, the output set for the outputs being weakly disposable
and the desirable outputs being strongly disposable is
n n
PðX Þ = Yj λj X ij ≤ X i , i = 1, . . . , m, τ λj Y rj = Y r , r = 1, . . . , b,
j=1 j=1
n
τ λj Y rj ≥ Y r , r = b þ 1, . . . , s, 0 ≤ τ ≤ 1, λj ≥ 0, j = 1, . . . , ng ð6:9Þ
j=1
By defining μj = τλj, P(X) can also be expressed as
n n
PðX Þ = Yj μj X ij ≤ τX i , i = 1, . . . , m, μj Y rj = Y r , r = 1, . . . , b,
j=1 j=1
n
μj Y rj ≥ Y, r = b þ 1, . . . , s, 0 ≤ τ ≤ 1, λj ≥ 0, j = 1, . . . , n
j=1
This subsection introduces two models: hyperbolic and directional distance.
6.2.3.1 Hyperbolic Model
To increase the desirable outputs and decrease the undesirable outputs at the same
rate, Färe et al. (1989) proposed a hyperbolic output model to measure efficiency:
max: φ
n
s:t: μj X ij ≤ τX i0 , i = 1, . . . , m
j=1
n
μj Y rj = Y r0 =φ, r = 1, . . . , b,
j=1 ð6:10Þ
n
μj Y rj ≥ φY r0 , r = b þ 1, . . . , s
j=1
0≤τ≤1
λj ≥ 0, j = 1, . . . , n
Note that in Färe et al. (1989) the τ associated with Xi0 in the first constraint is
mistakenly missing. However, this will not affect the solution because setting τ = 1
is feasible to Model (6.10), and it gives the largest feasible region. For the five
DMUs in Fig. 6.6, with the data shown in Table 6.1, Model (6.10) shows that DMUs
B and C are efficient, and DMUs A, D, and E are inefficient. The fifth column of
Table 6.1 shows the efficiencies of the five DMUs. Using DMU A to explain the
efficiency measure φ, the projection point is A, and the ratio of the coordinates of A
and A is φ, i.e. Y =Y 2A = φ and Y =Y 1A = 1=φ.
2A 1A
Table 6.1 Data and effi- DMU X Y1 (YU) Y2 (YD) φ δ

ciency measures of the weak
A 1 2 1 1.8257 0.5385
disposability approach
B 1 3 5 1 0
C 1 4 6 1 0
D 1 6 4 1.5 0.4
E 1 7 2 2.4152 0.7073
Note that the efficiency frontier is the line segments OBC, instead of CDE,
because the undesirable output Y1 can be reduced only accompanied by the reduction
of the desirable output Y2. An interesting case is that DMU A is inefficient, while it is
Pareto efficient. The reason is that although it produces smaller amounts of both the
desirable and undesirable outputs as compared to DMU B, they are not reducing at
the same rate as that seen in DMU B. From this point of view, the weak disposability
model is too stringent to be able to identify Pareto efficient DMUs.
6.2.3.2 Directional Distance Model
Model (6.10) is nonlinear, which may cause difficulties in the efficiency measure-
ment when there are a large number of DMUs. Based on the idea of directional
distance function, Färe and Grosskopf (2004) proposed a directional distance mea-
sure under the assumption of weak disposability of outputs:
max: δ
n
s:t: μj X ij ≤ τX i0 , i = 1, . . . , m
j=1
n
μj Y rj = Y r0 - δgU
r , r = 1, . . . , b,
j=1 ð6:11Þ
n
μj Y rj ≥ Y r0 þ δgD
r , r = b þ 1, . . . , s
j=1
0≤τ≤1
λj ≥ 0, j = 1, . . . , n
where g = gU r , gr is a direction vector. Again, the τ associated with Xi0 in the first
D
constraint is mistakenly missing in Färe and Grosskopf (2004), although it will not
affect the optimal solution. A DMU is efficient if δ = 0. Inefficient DMUs have
positive δ, and less efficient DMUs have larger values of δ. One merit of this model
is that it is linear. Another is that it is more general, in that various weights can be
assigned to different outputs by using different g. For the example in Table 6.1, if Y0
is assigned to g, then the direction for finding the projection point is d = (-Y10, Y20),
where the undesirable output Y1 will be reduced by Y10/Y20 for each unit reduction in
the desirable output Y2.
By using g = Y0, the efficiency measures of the five DMUs are calculated, with
the scores shown in the last column of Table 6.1. Using DMU A to explain,
g = (2, 1), the projection point is A, which is equal to A + δ(-2, 1). The results
obtained from this model are consistent with those obtained from the hyperbolic
model, in that both models fail to identify DMU A as efficient, and the orders of the
three inefficient DMUs are the same.
6.2.3.3 Variable-Reduction Model
In Eq. (6.9) the reduction factor τ is assumed to be the same for all DMUs.
Kuosmanen (2005) relaxed this assumption to allow each DMU to have its own
factor τj, and defined the output set as:
n n
PðX Þ = Yj λj X ij ≤ X i , i = 1, . . . , m, τj λj Y rj = Y r , r = 1, . . . , b,
j=1 j=1
n
τj λj Y rj ≥ Y r , r = b þ 1, . . . , s, 0 ≤ τj ≤ 1, λj ≥ 0, j = 1, . . . , n
j=1
For example, Model (6.11) based on the variable-reduction PðX Þ is:
max: δ
n
s:t: λj X ij ≤ X i0 , i = 1, . . . , m
j=1
n
τj λj Y rj = Y r0 - δgU
r , r = 1, . . . , b,
j=1
n
τj λj Y rj ≥ Y r0 þ δgD
r , r = b þ 1, . . . , s
j=1
0 ≤ τj ≤ 1
λj ≥ 0, j = 1, . . . , n
By setting τjλj = μj and (1 - τj)λj = θj, this model can be transformed into the
following linear model:
max: δ
n
s:t: μj þ θj X ij ≤ X i0 , i = 1, . . . , m
j=1
n
μj Y rj = Y r0 - δgU
r , r = 1, . . . , b, ð6:12Þ
j=1
n
μj Y rj ≥ Y r0 þ δgD
r , r = b þ 1, . . . , s
j=1
μj , θj ≥ 0, j = 1, . . . , n
Allowing the reduction factor to vary increases the flexibility in seeking the
maximum objective value. The variable-reduction model thus produces greater δ
values than the constant-reduction model. However, the variable-reduction model
produces the same δ values as those obtained from the constant-reduction model for
the data in Table 6.1.
6.2.4 Slacks-Based Approach
The slacks-based measure (SBM) models have been widely applied to measure
efficiencies. They can also be used to measure efficiency when there are undesirable
factors.
When smaller and larger amounts of desirable inputs and outputs are preferred,
respectively, larger and smaller amounts of undesirable inputs and outputs are also
preferred. In this regard, the strong disposability can be extended to include unde-
sirable factors, so that if (XD, XU, YD, YU) 2 T, then (XD′, XU′, YD′, YU′) 2 T, for XD′
≥ XD, XU′ ≤ XU, YD′ ≤ YD, YU′ ≥ YU, where T is the production possibility set. For a
set of n observations X D j , X j , Y j , Y j , j = 1, . . ., n, the corresponding production
U D U
possibility set is
n n
T= XD , XU , Y D , Y U j λj X D
j ≤X ,
D
λj X U
j ≥X ,
U
j=1 j=1
n n
λj Y D
j ≥Y ,
D
λj Y U
j ≤ Y , λj ≥ 0, j = 1, . . . , n
U
j=1 j=1
Let the first a inputs be undesirable and the remaining m-a be desirable;
similarly, let the first b outputs be undesirable and the remaining s-b be desirable.
Liu et al. (2010) proposed the following slacks-based model to measure efficiency:
1 m b
1- sD - =X i0 þ sU - =Y r0
ð m - aÞ þ b i = aþ1 i r=1 r
R0 = min:
1 a s
1þ sUþ =X i0 þ sDþ =Y r0
a þ ð s - bÞ i=1 i r = bþ1 r
n
s:t: λj X ij - sUþ
i = X i0 , i = 1, . . . , a
j=1
n
-
λj X ij þ sD
i = X i0 , i = a þ 1, . . . , m ð6:13Þ
j=1
n
-
λj Y rj þ sU
r = Y r0 , r = 1, . . . , b
j=1
n
λj Y rj - sDþ
r = Y r0 , r = b þ 1, . . . , s
j=1
- Uþ Dþ U -
λj , sD
i , si , sr , sr ≥ 0, 8j, i, r:
The DMU is efficient if R0 = 1 and is inefficient if R0 < 1. The objective function

is nonlinear; however, it can be linearized, as demonstrated in Chap. 4.
Several approaches are introduced in this section. In general, however, it is not
appropriate to say one is better than another, and the choice of which one to use must
be made depending on the nature of the applications.
The non-discretionary factor model proposed by Banker and Morey (1986) may
overestimate the efficiency if the non-discretionary factors are correlated with the
efficiency. Ray (1991) applied a regression analysis, using the non-discretionary
factors as the independent variables, to distinguish the efficiency from the
non-discretionary effects. Ruggiero (1996) tackled this problem by constructing
shifted frontiers for different levels of the non-discretionary factors. Ruggiero
(1998) extended the ideas of the two studies to develop a three-stage model. Several
simulation studies have been conducted to investigate the effect of the
non-discretionary factors on efficiencies (Muñiz et al., 2006; Syrjanen, 2004; Yu,
1998). In addition to the radial measures, slacks-based measures have also been
proposed (Jahanshahloo et al., 2010; Muñiz, 2002).
The undesirable factors are important issues in environmental and energy studies
and have been widely discussed in the related literature (Leleu, 2013; Sahoo et al.,
2011; Song et al., 2012; Sueyoshi & Wang, 2014; Wang et al., 2012; You & Yan,
2011; Zhou et al., 2008). Lovell and Pastor (1995), Pastor (1996), Dyson et al.
(2001), Cherchye et al. (2011), and Zanella et al. (2015) also discuss the data
transformation approach, while Kuosmanen (2005), Podinovski and Kuosmanen
(2011), Kuosmanen and Matin (2011), and Kao and Hwang (2023) have some
References 133
thoughts on the variable-reduction weak disposability approach. Chung et al. (1997)

is probably the first study of directional distance functions, and other works that have
examined these include Seiford and Zhu (2005), Halkos and Tzeremes (2013),
Hwang et al. (2013), Chen et al. (2015), and Diabat et al. (2015). The slacks-based
model introduced in this chapter is a general one, and there are other, more specific
models, such as Zhou et al. (2006), Lozano and Gutierrez (2011), Zhang and Choi
(2013), Li and Wang (2014), and Bi et al. (2015).
References
Ali, A. I., & Seiford, L. M. (1990). Translation invariance in data envelopment analysis. Operations
Research Letters, 9, 403–405.
Banker, R. D., & Morey, R. C. (1986). Efficiency analysis for exogenously fixed inputs and outputs.
Operations Research, 34, 513–521.
Bi, G. B., Luo, Y., Ding, J. J., & Liang, L. (2015). Environmental performance analysis of Chinese
industry from a slacks-based perspective. Annals of Operations Research, 228, 65–80.
Chen, P. C., Yu, M. M., Chang, C. C., Hsu, S. H., & Managi, S. (2015). The enhanced Russell-
based directional distance measure with undesirable outputs: Numerical example considering
CO2 emissions. Omega, 53, 30–40.
Cherchye, L., Moesen, W., Rogge, N., & Van Puyenbroeck, T. (2011). Constructing composite
indicators with imprecise data: A proposal. Expert Systems with Applications, 38, 10940–10949.
Chung, Y. H., Färe, R., & Grosskopf, S. (1997). Productivity and undesirable outputs: A directional
distance function approach. Journal of Environmental Management, 51(3), 229–240.
Diabat, A., Shetty, U., & Pakkala, T. P. M. (2015). Improved efficiency measures through
directional distance formulation of data envelopment analysis. Annals of Operations Research,
229, 325–346.
Dyson, R. G., Allen, R., Camanho, A. S., Podinovski, V. V., Sarrico, C. S., & Shale, E. A. (2001).
Pitfalls and protocols in DEA. European Journal of Operational Research, 132, 245–259.
Färe, R., & Grosskopf, S. (2004). Modeling undesirable factors in efficiency evaluation: Comment.
Färe, R., Grosskopf, S., Lovell, C. A. K., & Pasurka, C. (1989). Multilateral productivity compar-
isons when some outputs are undesirable: A nonparametric approach. The Review of Economics
and Statistics, 71, 90–98.
Golany, B., & Roll, Y. (1989). An application procedure for DEA. Omega, 17, 237–250.
Halkos, G. E., & Tzeremes, N. G. (2013). A conditional directional distance function approach for
measuring regional environmental efficiency: Evidence from UK regions. European Journal of
Hwang, S. N., Chen, C., Chen, Y., Lee, H. S., & Shen, P. D. (2013). Sustainable design
performance evaluation with applications in the automobile industry: Focusing on inefficiency
by undesirable factors. Omega, 41, 553–558.
Jahanshahloo, G. R., Lotfi, F. H., Shoja, N., Abri, A. G., Jelodar, M. F., & Firouzabadi, K. J. (2010).
Proposing a new model on data envelopment analysis by considering non discretionary factors
and a review on previous models. Mathematical & Computational Applications, 15, 344–353.
Kao, C., & Hwang, S. N. (2023). Separating the effect of undesirable outputs generation from the
inefficiency of desirable outputs production in efficiency measurement. European Journal of
Operational Research (accepted).
Koopmans, T. C. (1951). Analysis of production as an efficient combination of activities. In T. C.
Koopmans (Ed.), Activity analysis of production and allocation (pp. 33–97). Wiley.
Kuosmanen, T. (2005). Weak disposability in nonparametric production analysis with undesirable

outputs. American Journal of Agricultural Economics, 87, 1077–1082.
Kuosmanen, T., & Matin, R. K. (2011). Duality of weakly disposable technology. Omega, 39,
504–512.
Leleu, H. (2013). Shadow pricing of undesirable outputs in nonparametric analysis. European
Li, M. Q., & Wang, Q. (2014). International environmental efficiency differences and their
determinants. Energy, 78, 411–420.
Liu, W. B., Meng, W., Li, X. X., & Zhang, D. Q. (2010). DEA models with undesirable inputs and
outputs. Annals of Operations Research, 173, 177–194.
Lovell, C. A. K., & Pastor, J. T. (1995). Units invariant and translation invariant DEA models.
Operations Research Letters, 18, 147–151.
Lozano, S., & Gutierrez, E. (2011). Slacks-based measure of efficiency of airports with airplanes
delays as undesirable outputs. Computers & Operations Research, 38, 131–139.
Muñiz, M. (2002). Separating managerial inefficiency and external conditions in data envelopment
Muñiz, M., Paradi, J., Ruggiero, J., & Yang, Z. J. (2006). Evaluating alternative DEA models used
to control for non-discretionary inputs. Computers & Operations Research, 33, 1173–1183.
Pastor, J. T. (1996). Translation invariance in DEA: A generalization. Annals of Operations
Research, 66, 93–102.
Pittman, R. W. (1983). Multilateral productivity comparisons with undesirable outputs. Economic
Journal, 93(372), 883–891.
Podinovski, V. V., & Kuosmanen, T. (2011). Modelling weak disposability in data envelopment
analysis under relaxed convexity assumptions. European Journal of Operational Research,
211, 577–585.
Ray, S. C. (1991). Resource-use efficiency in public-schools—A study of Connecticut data.
Management Science, 37, 1620–1628.
Ruggiero, J. (1996). On the measurement of technical efficiency in the public sector. European
Ruggiero, J. (1998). Non-discretionary inputs in data envelopment analysis. European Journal of
Ruggiero, J. (2004). Performance evaluation when non-discretionary factors correlate with techni-
cal efficiency. European Journal of Operational Research, 159, 250–257.
Sahoo, B. K., Luptacik, M., & Mahlberg, B. (2011). Alternative measures of environmental
technology structure in DEA: An application. European Journal of Operational Research,
215, 750–762.
Scheel, H. (2001). Undesirable outputs in efficiency valuations. European Journal of Operational
Research, 132, 400–410.
Seiford, L. M., & Zhu, J. (2002). Modeling undesirable factors in efficiency evaluation. European
Seiford, L. M., & Zhu, J. (2005). A response to comments on modeling undesirable factors in
efficiency evaluation. European Journal of Operational Research, 161, 579–581.
Song, M., An, Q., Zhang, W., Wang, Z., & Wu, J. (2012). Environmental efficiency evaluation
based on data envelopment analysis: A review. Renewable and Sustainable Energy Reviews, 16,
4465–4469.
Sueyoshi, T., & Wang, D. (2014). Radial and non-radial approaches for environmental assessment
by data envelopment analysis: Corporate sustainability and effective investment for technology
innovation. Energy Economics, 45, 537–551.
Syrjanen, M. J. (2004). Non-discretionary and discretionary factors and scale in data envelopment
Wang, K., Wei, Y. M., & Zhang, X. (2012). A comparative analysis of China’s regional energy and
emission performance: Which is the better way to deal with undesirable outputs? Energy Policy,
46, 574–584.
References 135
You, S., & Yan, H. (2011). A new approach in modeling undesirable output in DEA model. Journal
of the Operational Research Society, 62, 2146–2156.
Yu, C. (1998). The effects of exogenous variables in efficiency measurement—A Monte Carlo
study. European Journal of Operational Research, 105, 569–580.
Zanella, A., Camanho, A. S., & Dias, T. G. (2015). Undesirable outputs and weighting schemes in
composite indicators based on data envelopment analysis. European Journal of Operational
Research, 245, 517–530.
Zhang, N., & Choi, Y. (2013). Environmental energy efficiency of China’s regional economies: A
non-oriented slacks-based measure analysis. Social Science Journal, 50, 225–234.
Zhou, P., Ang, B. W., & Poh, K. L. (2006). Slacks-based efficiency measures for modeling
environmental performance. Ecological Economics, 60, 111–118.
Zhou, P., Ang, B. W., & Poh, K. L. (2008). A survey of data envelopment analysis in energy and
environmental studies. European Journal of Operational Research, 189, 1–18.
Kao, C. & Hwang, S. N. (2023). Separating the effect of undesirable outputs generation from the
inefficiency of desirable outputs production in efficiency measurement. European Journal of
Operational Research (accepted)
Chapter 7
Special Types of Data
In the production process physical inputs are consumed to produce physical outputs.
All the data are therefore positive real numbers, and the conventional DEA models
are based on this requirement. However, in many applications the inputs and outputs
may not be physical factors, and thus the data are not necessarily positive real
numbers. For example, if profit is used as an output, then it can be negative, although
the physical quantity of output is always positive. Another situation is that the
physical quantities of the inputs consumed by the DMUs are ordered in ranks, or
the abilities of different persons (considered as DMUs) can only be ranked, without
real term measures. In these cases the data only reflects the relative differences
among the DMUs, instead of the absolute differences represented by the conven-
tional measures. A similar case is that the service level provided by a DMU cannot
be evaluated by any measures and is only subjectively expressed by linguistic terms,
such as excellent, good, acceptable, and unacceptable. Data of these three types are
referred to as negative, ordinal, and qualitative, and Sections 7.1, 7.2, and 7.3 will
introduce models for handling these.
Even if the inputs and outputs are physical factors, the data can still be different
from the conventional precise values. For example, due to the uncertain environment
or imprecision in measurement, the data is not a precise value. Sections 7.4, 7.5, and
7.6 introduce three ideas for measuring efficiencies under these circumstances. The
first is to treat it as a random variable with a specified probability distribution. The
second is to treat it as an interval data, covering a range of values. The third is to use
fuzzy numbers to represent the imprecise data. These three types of data are referred
to as stochastic, interval, and fuzzy.
https://doi.org/10.1007/978-3-031-27593-7_7
138 7 Special Types of Data
7.1 Negative Data
Although the input and output data in normal conditions are positive, there are cases
that the data may be negative, such as when dealing with economic growth, profit
earned, and changes in the birth rate. The conventional DEA models that require all
data to be positive must be modified to accommodate negative values.
If an output factor has negative values for all DMUs, then one simple way is to
change it to an input factor, with the minus sign deleted. The rationale is that more
negative values, as an output, indicate worse performance, while larger positive
values, as an input, also indicate worse performance. The negative output can thus be
treated as a positive input. Similarly, if an input factor has negative values for all
DMUs, then it can be changed to an output, with the sign of the data reversed. This
input–output exchange method, however, can be used only when the values of all
DMUs are negative. Moreover, treating a negative output (or input) as a positive
input (or output) does not reflect the true production process.
One more formal way for handling negative values is to use the directional
distance measure introduced in Model (3.19) of Chap. 3, which is rewritten below
to maintain continuity:
max:η
n
s:t: λj X i j ≤ X i0 - ηf i , i = 1, . . . , m
j=1
n
λj Y r j ≥ Y r0 þ ηgr , r = 1, . . . , s ð7:1Þ
j=1
n
λj = 1
j=1
λj ≥ 0, j = 1, . . . , n
where d = (-f, g) is the direction for the DMU moving toward the frontier in
measuring inefficiency. The DMU is efficient if η = 0, and is inefficient if
η > 0, where a larger η implies a less efficient DMU. In the setting of Model
(7.1), not all data for a factor need to be negative.
Consider six DMUs using one-input X to produce one output Y, with the data
shown in columns two and three of Table 7.1, and the six DMUs depicted in Fig. 7.1.
The production frontier is the line segments ABC. Any point on the frontier that
dominates the DMU to be evaluated can be selected as the projection point, and
Cheng et al. (2013) suggested using dA = (-| X0| , | Y0| ) to find this. Using the
absolute values of Xi0 and Yr0 to generate the direction ensures that the frontier
will definitely be reached. Column four of Table 7.1 shows the directions for
measuring inefficiencies of the six DMUs, and column five shows the distance
measures calculated from Model (7.1). DMUs A, B, and C lie on the frontier and
7.1 Negative Data 139
Table 7.1 Data and direc- Absolute direction Range direction

tional distance measures for
DMU X Y Direction dA η Direction dR η
the negative-data example
A -2 -2 (-2, 2) 0 (0, 4) 0
B -1 1 (-1, 1) 0 (-1, 1) 0
C 3 2 (-3, 2) 0 (-5, 0) 0
D -1 -3 (-1, 3) 2/3 (-1, 5) 1/2
E 0 -1 (0, 1) 9/4 (-2, 3) 5/9
F 2 1 (-2, 1) 1/2 (-4, 1) 3/8
Fig. 7.1 Directional Y

distance efficiency
measurement for the
negative-data example C
2 d FA
d A d FR
E
B 1 F
d ER
X
–3 – 2 d DR – 1 O 1 2 3
d DA
E –1
A –2
D –3
are thus efficient, with η = 0. The other three DMUs are inefficient, where E performs
better than F, and F performs better than D.
Although Model (7.1) is able to identify efficient and inefficient DMUs, it was
noted in Kerstens and van de Woestyne (2014) that the input-oriented efficiency
measure might be negative. Moreover, Model (7.1) is not unit invariant. If different
units are used to measure the input and output factors, then the results will be
different. To solve this problem, Portela et al. (2004) proposed a range directional
measure that uses the largest possible step that the DMU can move as the direction
vector, so that the distance η will be less than or equal to one, with 1-η defined as the
efficiency. Specifically, the direction is defined as:
f i = X i0 - min j = 1,...,n fX ij g, i = 1, . . . , m
ð7:2Þ
gr = max j = 1,...,n fY rj g - Y r0 , r = 1, . . . , s:
With this setting, the efficiency 1-η will lie between 0 and 1, and Model (7.1)
will be unit invariant. Moreover, under variable returns to scale, Model (7.1) is also
translation invariant. The last two columns of Table 7.1 show the directions dR = (-
f, g) obtained from Expression (7.2), and the corresponding inefficiencies calculated
from Model (7.1). The three efficient and inefficient DMUs are identified correctly.
However, the rankings of the three inefficient DMUs are different from those
measured from the absolute-direction method. While the rankings of DMUs D, E,
and F are 6, 4, and 5, respectively, from the absolute-direction method, they are 5, 6,
and 4, respectively, from the range-direction method. This difference is actually a
characteristic of the directional distance measure, in that different directions yield
different measures for inefficient DMUs.
While the range-direction method is able to obtain reasonable efficiency mea-
sures, it was found in Kao (2020) that the calculated efficiencies are biased, and a
generalized radial model requiring the aggregate output value sr = 1 ur Y rj and the
aggregate input value m i = 1 vi X ij to be positive was proposed to produce reliable
measures. The idea is that if some outputs of negative values are produced, then they
must be accompanied with some outputs of positive values at the same time, and the
positive values must outweigh the negative values to make the production process
worth undertaking. Similarly, if some negative inputs occur in a production process,
then some positive inputs must be consumed at the same time such that the positive
values outweigh the negative values. Otherwise, this production process is generat-
ing resources while producing outputs. The corresponding BCC output is:
m
1
= min:v0 þ vi X ik
Ek i=1
s
s:t: ur Y rk = 1
r=1
m s
v0 þ vi X i j - ur Y r j ≥ 0, j = 1, . . . , n
i=1 r=1
ð7:3Þ
m
vi X i j ≥ 0, j = 1, . . . , n
i=1
s
ur Y r j ≥ 0, j = 1, . . . , n
r=1
ur ,vi ≥ ε, r = 1, . . . ,s, i = 1, . . . , m
The assessed DMU is allowed to select the most favorable multipliers ur and vi to
make the aggregate input and aggregate output values positive. When all observa-
tions are positive, the constraints m i = 1 vi X ij ≥ 0 and
s
r = 1 ur Y rj ≥ 0 become redun-
dant, and this model boils down to the conventional BCC Model.
m s
The constraints i = 1 vi X ij ≥ 0 and r = 1 ur Y rj ≥ 0 in Model (7.3) limit the
production process to those that the aggregate input and aggregate output values
are positive. Model (7.3) is thus not applicable to general cases of negative data. For
7.2 Ordinal Data 141
example, when there is only one input or one output with negative data, then this
model is not able to produce efficiency scores for the DMUs to be evaluated. In this
case, the more general radial model developed in Kao (2022) can be applied, which
the interested readers can refer to.
7.2 Ordinal Data
The data used in measuring efficiency is normally of a ratio type, in statistical terms,
which is able to distinguish the exact difference between two values. In contrast,
there is also ordinal data which only shows the order of the values, and the difference
between two values is not really known. For example, on a five-point scale the
difference between a 4 and 5 is not necessarily the same as that between a 1 and 2.
The basic concept for measuring the efficiency of a set of DMUs with ordinal data
is still “the most favorable condition.” Let the last output Ys be ordinal, and the input
can be handled similarly. We want to assign positive real values to different ordinal
numbers that will produce the largest possible efficiency score for the DMU being
evaluated. Let yk, k = 1, . . ., n, be the value assigned to ordinal number k. The
precedence requirement of yk > yk + 1 must be satisfied. Since the units of y do not
affect the measured efficiency, one can set y1 = 1. The strong inequality of yk > yk + 1
can be expressed by the weak inequality of yk ≥ yk + 1 + δ so that mathematical
computations can be carried out, where δ is a small number that is large enough to
differentiate the order of yk and yk + 1. Cooper et al. (1999) proposed the following
model to measure the efficiency of a DMU when output s is of ordinal data:
s-1
E 0 = max: ur Y r0 þ us y½0 - u0
r=1
m
s:t: vi X i0 = 1
i=1
s-1 m ð7:4Þ
ur Y r j þ us y½j - u0 - vi X i j ≤ 0, j = 1, . . . , n
r=1 i=1
y1 = 1, yk ≥ ykþ1 þ δ, k = 1, . . . , n - 1
ur ,vi ,yk ≥ ε, r = 1, . . . , s, i = 1, . . . , m, k = 1, . . . , n
u0 unrestricted in sign,
where the subscript [j] indicates the order of DMU j in Ys.

Model (7.4) is nonlinear due to the nonlinear term usy[j]. By applying a variable
substitution of w[j] = usy[j], Model (7.4) is transformed into the following linear
program:
s-1
E 0 = max: ur Y r0 þ w½0 - u0
r=1
m
s:t: vi X i0 = 1
i=1
s-1 m ð7:5Þ
ur Y r j þ w½j - u0 - vi X i j ≤ 0, j = 1, . . . , n
r=1 i=1
w1 = us , wk ≥ wkþ1 þ us δ, k = 1, . . . , n - 1
ur ,vi ,wk ≥ ε, r = 1, . . . , s, i = 1, . . . , m, k = 1, . . . , n
Although this model is able to measure the efficiency for DMUs with ordinal
data, the measured efficiency scores are superficial, in that the difference between
two scores is the artifact effect of δ. For example, suppose there are four DMUs,
A, B, C, and D, using the same amount of one unit of input X to produce the ordinal
output Y, with YA = 3, YB = 1, YC = 4, and YD = 2. By applying Model (7.4), the
efficiencies obtained are EA = 1 - 2δ, EB = 1, EC = 1 - 3δ, and ED = 1 - δ, which
do not seem to reflect the true efficiencies appropriately. This is obviously due to
using the weak inequality of yk ≥ yk + 1 + δ to represent the strong inequality of
yk > yk + 1. More suitable ways of transforming the strong inequality into weak
inequality are thus needed, as well as other methods for handling ordinal data.
7.3 Qualitative Data
Another type of data that can also be considered as ordinal is qualitative data. This is
often described using linguistic terms or symbols, which only show the precedence,
rather than the difference between the data. For example, the symbols A, B, C, D, and
F are generally used in grading students’ work, with the preceding ones better than
the succeeding ones. However, how much better the former are than the latter is not
clear. Similarly, the education level can be doctor, master, and baccalaureate, where
the former is higher than the latter, although the exact differences among them are
not known. In behavioral science studies customer satisfaction regarding a service
may be surveyed, and verbal descriptions such as very much satisfied, satisfied,
acceptable, and unacceptable are often used. How to express such qualitative data in
a form that can be computed is a challenging task.
One way of dealing with this problem is to use variables to represent the values
that should be assigned to the qualitative data, similar to the approach used with
ordinal data. Suppose the last input is expressed by the qualitative data of very large,
large, medium, small, and very small. They can then be represented by
1 = x1 > x2 > x3 > x4 > x5. The strong inequality of xk > xk + 1 is replaced with
the weak inequality of xk ≥ xk + 1 + δ. When the variable xk is multiplied by the
7.3 Qualitative Data 143
multiplier vm, the result is a nonlinear term of vmxk. Similar to the case of substituting
the ordinal data by a new variable wk, with the constraints of 1 = x1 > x2 > x3 > x4 > x5
replaced with w1 = vm, wk ≥ wk + 1 + vmδ, k = 1, . . ., 4, the problem then is
computationally operable.
Let x[j] represent the effort that DMU j devotes to a job. The input BCC model for
measuring efficiency is
s
E 0 = max: ur Y r0 - u0
r=1
m-1
s:t: vi X i0 þ w½0 = 1
i=1
s m-1 ð7:6Þ
ur Y r j - u 0 - vi X i j þ w½j ≤ 0, j = 1, . . . , n
r=1 i=1
w1 = vm , wk ≥ wkþ1 þ vm δ, k = 1, . . . , n - 1
ur ,vi ,wk ≥ ε, r = 1, . . . , s, i = 1, . . . , m, k = 1, . . . , n
Different from Model (7.5) for ordinal data, in which each DMU has a specific
order, several DMUs can have the same level in Model (7.6) for qualitative data. The
output model can be formulated similarly.
Another way of handling the qualitative data is to treat different levels as
categorical variables (Banker and Morey 1986), so that DMUs of the lowest level
are only compared with DMUs of the same level. In contrast, DMUs of higher levels
are compared with not only DMUs of the same level, but also those of lower levels.
The rationale is that, other things being equal, DMUs of a specific level should
perform better than those of lower levels. If there is a DMU of a lower level that
performs better than this DMU, then this DMU is inefficient.
Without loss of generality, let the last input Xm be a categorical variable. The
production possibility set in this case is given by:
n n
T = ðX, Y Þ j λj Y rj ≥ Y r , r = 1, . . . , s, λj X ij ≤ X i , i = 1, . . . , m - 1,
j=1 j=1
n
λj = 1, λj ≥ 0, j = 1, . . . , n, X mj ≤ X m for λj > 0 ð7:7Þ
j=1
The conventional constraint of nj = 1 λj X ij ≤ X m0 for the categorical variable in

measuring the efficiency of a DMU is replaced with Xmj ≤ Xm0 for those DMUs used
in the convex combination in expressing this DMU. Since which λj will be nonzero
is not known beforehand, this conceptual form of constraint must be expressed in a
form that can be implemented computationally.
Table 7.2 Data and efficiency measures for the qualitative-data example
DMU X Y1 Y2 Y3 Categorical Efficiency Value-assigning Efficiency
A 1 3 3 Bad 1 1/1.00022
B 1 2 3 Fair 1/(1 + ε) 1/1.00031
C 1 6 2 Fair 1 1
D 1 5 2 Good 11/12 1
E 1 4 4 Good 1 1
F 1 1 5 Excellent 1 1
The constraints in (7.7) essentially state that the DMU being evaluated will be
compared only with those of the same or lower levels. Suppose there are l levels for
ðk Þ
the categorical variable. Let Bmj = 0 or 1 be a binary constant corresponding to level
ð1Þ ð2Þ ðk Þ ðkþ1Þ
k. If DMU j has a level of k, then it is expressed as Bmj =Bmj =. . .=Bmj =1 and Bmj
ðkþ2Þ ðlÞ
=Bmj =. . .=Bmj = 0. With this notation, the conceptual constraint of Xmj ≤ Xm0 for
n ðk Þ ðk Þ
λj > 0, j = 1, . . ., n, can be replaced by j = 1 λj Bmj ≤ Bm0 , k = 1, . . ., l, and the
corresponding output BCC model is:
max:φ
n
s:t: λj X i j ≤ X i0 , i = 1, . . . ,m - 1
j=1
n
λj Y r j ≥ φY r0 , r = 1, . . . , s
j=1
n ð7:8Þ
λj = 1
j=1
n
ðkÞ ðkÞ
λj Bm j ≤ Bm0 , k = 1, . . . , l
j=1
λj ≥ 0, j = 1, . . . , n:
For example, if there are four levels, and the DMU being evaluated has a level of
ð1Þ ð2Þ
three, then the corresponding constraints will be nj = 1 λj Bmj ≤ 1, n
j = 1 λj Bmj ≤ 1,
ð3Þ ð4Þ
n
≤ 1, and nj = 1 λj Bmj ≤ 0: With this set of constraints λj will be positive
j = 1 λj Bmj
only for those DMUs of the same or lower levels as the DMU being evaluated, which
implies that this DMU will only be compared with those of the same or lower levels.
Consider six DMUs using the same amount of input X to produce different
combinations of three outputs Y1, Y2, and Y3, where Y3 is a categorical variable,
with the data shown in Table 7.2, and the DMUs depicted in Fig. 7.2. Since DMU
A is the only DMU that is of the lowest level, Bad, the production frontier for level
Bad is T1AT1', and DMU A is efficient. The production frontier for level Fair is
T1ACT2', which is constructed from the DMUs with levels of Fair and Bad. In this
7.3 Qualitative Data 145
Fig. 7.2 Frontiers for the Y2

categorical-data example
6
F
5 T4
4 T3 E
B A
3 T1
D*
2 D C
,
T1
1
,
T2
O 1 2 3 4 5 6
Y1
case, DMUs B and C, with level Fair, are weakly and strongly efficient, respectively.
By the same token, the frontiers for levels Good and Excellent are T3ECT2' and
T4FECT2', respectively, where only DMU D is inefficient, with an efficiency of OD/
OD = 11/12..
If the method of assigning values to different levels similar to that discussed in
Model (7.6) is used, but in an output-oriented form with ε = 10-4 and δ = 10-5, then
the resulting efficiencies are those shown in the last column of Table 7.2. The results
are quite different from those measured from the categorical variable method. All
DMUs are either strongly or weakly efficient, with two DMUs, A and B, being
identified as inefficient. However, the positions of the six DMUs on the Y1-Y2 plane
indicate that the results calculated from the categorical method seem to be more
reasonable. The high efficiencies calculated from the value-assigning method are
due to assigning sufficiently close values to variables representing different levels,
for example, y4 = 1, y3 = 1 - δ, y2 = 1 - 2δ, and y1 = 1 - 3δ. The categorical
variable method is not without weaknesses, either. When the number of the DMUs
of the lowest level is not large enough, the calculated efficiency will be overstated.
Since qualitative data is usually described by linguistic terms, and fuzzy numbers
are suitable to express these, the fuzzy number method has also been used to measure
the efficiency of DMUs with qualitative data. However, the fuzzy numbers are
subjectively assigned to the linguistic terms, rather than objectively reflected from
the qualitative data itself, which does not accord with the spirit of DEA. Kao and Lin
(2011) studied a specific problem and were able to objectively obtain four fuzzy
numbers to represent the four levels of a qualitative variable. The idea of fuzzy
numbers will be discussed in more detail in Sect. 7.6.
7.4 Stochastic Data
Many studies have discussed the issue of stochastic observations, and one major
branch of these adopts the parametric approach of stochastic frontier analysis (SFA).
Since this book concentrates on the nonparametric approach of data envelopment
analysis, the SFA technique is not discussed. Interested readers can refer to, for
example, the book by Kumbhaker and Lovell (2000).
In measuring the past performance of a set of DMUs over a period of time, the
average data of a specific unit of time, e.g. a year, is generally used. While this gives
a general idea of the performance in a fast and easy way, the results may not reflect
the true performance of the DMUs, especially their rankings, because the data
fluctuates over time, and the average is just one of many possible circumstances.
Another situation is in predicting future performance, when historical data, as
modified by human judgment, is generally used. In these cases the data is not
deterministic, but rather stochastic, and described by probability distributions.
When data is stochastic, the resulting efficiency score will also be stochastic.
Treating the efficiency calculated from the average data as the performance of a
DMU will be misleading and make the decision maker over-confident with the
results.
Consider a very simple case of three DMUs, A, B, and C, using input X, at levels
of 1, 2, and 3, respectively, to produce output Y, of the amounts of 2, Uniform (2.5,
4.5), and 4, respectively, as depicted in Fig. 7.3, where Uniform (2.5, 4.5) indicates
that the output of DMU B, YB, is a random variable uniformly distributed over (2.5,
4.5). The conventional way of measuring efficiency in this case is to use the average
of 3.5 to represent the random output of YB, to obtain the frontier ABC, where all
three DMUs are efficient. When the stochastic nature of YB is considered, where YB is
described by the uniform distribution function of
Fig. 7.3 Frontiers for the Y

stochastic-data example
5
B
4 C
3
B
2 A
O 1 2 3 X
7.4 Stochastic Data 147
0, y ≤ 2:5
F Y B ðyÞ = ðy - 2:5Þ=2, 2:5 ≤ y ≤ 4:5
1, 4:5 ≤ y
the production frontier becomes stochastic. First, it is noted that DMU A is always
efficient, no matter what value of YB is. For DMUs B and C, different values of YB
result in different efficiencies, and there are two cases to discuss, YB 2 (2.5, 3) and
YB 2 (4, 4.5)..
For YB in the range of (2.5, 3), the production frontier is AC, which produces an
efficiency score of YB/3 for DMU B. When YB has a value greater than 3, the frontier
is ABC or ABB’, and DMU B becomes efficient. The cumulative density of the
random efficiency EB can then be derived as:
F EB ðeÞ = PrfEB ≤ eg = PrfY B =3 ≤ eg = PrfY B ≤ 3eg = ð3e - 2:5Þ=2:
For YB2(2.5, 3), EB is in the range of (2.5/3, 1), and the distribution of EB is:
0, e ≤ 5=6
F EB ðeÞ = ð3e - 2:5Þ=2, 5=6 ≤ e ≤ 1
1, 1≤e
Together with the probability that Pr{EB = 1} = Pr {YB ≥ 3} = 0.75, the

1
expected value of EB is thus E ½EB = 0:75 × ð1Þ þ 5=6 edF EB ðeÞ = 47=48, which is
smaller than the value of 1, calculated from the average data of YB = 3.5..
When YB is less than or equal to 4, the frontier is ABC or AC, and DMU C is
efficient. For YB 2 (4, 4.5), the frontier is ABB’, where DMU C is inefficient, with the
efficiency score of EC = 4/YB. The cumulative density of EC is
F EC ðeÞ = PrfE C ≤ eg = Prf4=Y B ≤ eg = PrfY B ≥ 4=eg = 1 - PrfY B ≤ 4=eg

= 1 - ð4=e - 2:5Þ=2 = 9=4 - 2=e:
The complete distribution function is:
0, e ≤ 8=9
F E C ð eÞ = 9=4 - 2=e, 8=9 ≤ e ≤ 1
1, 1≤e
Since Pr{EC = 1} = Pr {YB ≤ 4} = 0.75, the expected efficiency of DMU C is

1
E ½E C = 0:75 × ð1Þ þ 8=9 edF EC ðeÞ = 71=72, which is also smaller than 1, the value
calculated from the average data of YB = 3.5.
This example shows that using the expected data to calculate efficiencies results
in three efficient DMUs, while using the stochastic data obtains a result of only one
DMU which is always efficient, and the other two are efficient with a probability of
0.75. The decision maker will be over-confident with the results when the stochastic
data is replaced with the expected data.
The above analytical derivation is applicable only for very simple problems. For
general problems, Kao and Liu (2009) used a simulation to find the distribution of
the efficiency. The idea is simply to randomly select a set of input and output data
from the distributions of the stochastic data and apply the conventional DEA models
to measure the efficiency. By repeating this process a sufficiently large number of
times, an empirical distribution of the efficiency for each DMU is obtained. The
mean, variance, and the probability that the efficiency score will occur in an interval
can be calculated numerically. This method is able to produce accurate results as
long as the number of replications is large enough, and the only problem is the long
execution time involved.
In contrast to the numerical approach of simulation, there is an analytical
approach, based on the idea of chance-constrained programming of Charnes and
Cooper (1959, 1963). Consider a case where the inputs are observed with the
deterministic data Xij, while the outputs are to be predicted and are represented by
random variables yrj. The BCC input model in this case is:
s
e0 = max: ur yr0 - u0
r=1
m
s:t: vi X i0 = 1
i=1
ð7:9Þ
s m
ur y r j - u0 - vi X i j ≤ 0, j = 1, . . . , n
r=1 i=1
ur ,vi ≥ ε, r = 1, . . . , s, i = 1, . . . , m
There are two points to be noted in this model. One is that the objective function
is a random variable, which is computationally intractable. The other is that the
constraints are not guaranteed to be satisfied due to the random effect of yrj.
To make the model computationally operable, the objective function “max.
s
r = 1 ur yr0 - u0 ” is replaced with the expectation form of “max.
s
E u y
r = 1 r r0 - u 0 = max. sr = 1 ur μr0 - u0 ”, where μr0 is the mean of yr0, so
that the random efficiency e0 is a constant. The chance constraints are changed to the
following form and satisfied with a specified probability 1 - αj:
7.4 Stochastic Data 149
s m
Pr ur yrj - u0 ≤ vi X ij ≥ 1 - αj , j = 1, . . . , n
r=1 i=1
Let σpq be the covariance of ypj and yqj. The covariance of ypj and ypj is the
variance of ypj and is denoted as σ 2p . If we subtract the mean of sr = 1 ur μr0 - u0 on
both sides of the constraint and divide by the standard deviation of Vj =
s
var r = 1 ur yrj , we will get a random variable zj, which approximately follows
a standard normal deviation:
s s
r¼1 ur yrj - u0 - r¼1 ur μrj - u0
Pr zj ¼
Vj
m s
i¼1 vi X ij - r¼1 ur μrj - u0
≤ ≥ 1 - αj
Vj
This expression is equivalent to:
m s
i = 1 vi X ij - r = 1 ur μrj - u0
p ≥ Φ - 1 1 - αj
Vi
where Φ-1 is the inverse function of the standard normal distribution function. For
example, if αj = 0.025, then Φ-1(1 - αj) = 1.96.
The variance Vj is a quadratic function of the following form:
σ21 σ12 σ13 . . . σ1s u1

σ21 σ22 σ23 . . . σ2s u2
V j = ½u1 , u2 , . . . , us : : : ... : : ð7:10Þ
: : : ... : :
σs1 σs2 σs3 . . . σ2s us
The standard deviation V j is nonlinear. However, if we express yrj = μrj + brjς,

where ς is assumed to follow a Normal (0, σ 2) distribution, then yrj follows a Normal
μrj þ b2rj σ2 distribution, and Vj becomes the square of sr = 1 ur brj σ, as shown
below:
b21j σ2 b1j b2j σ12 b1j b3j σ13 . . . b1j bsj σ1s u1
b2j b1j σ21 b22j σ22 b2j b3j σ23 . . . b2j bsj σ2s u2
V j = ½u1 , u2 , . . . , us : : : ... : :
: : : ... : :
bsj b1j σs1 bsj b2j σs2 bsj b3j σs3 . . . bsj σs
2 2 us
s 2
= ur brj σ
r=1
With this representation, the standard deviation V j becomes linear, and Model
(7.9) can be expressed by the following linear programming model:
s
e0 = max: ur μr0 - u0
r=1
m
s:t: vi X i0 = 1
i=1
m s s
vi X i j - ur μr j - u0 ≥ ur br j σ Φ - 1 ð1 - αj Þ, j = 1, . . . , n
i=1 r=1 r=1
ur , vi ≥ ε, r = 1, . . . , s, i = 1, . . . , m
ð7:11Þ
In this model the mean μrj and standard deviation brjσ of yrj need to be specified.
If yrj has a beta distribution, then μrj is approximately equal to (OPrj + 4MLrj + PErj)/
6, where OPrj, MLrj, and PErj are the optimistic, most likely, and pessimistic
estimates of yrj, respectively. These measures can be obtained from past experience
or the opinions of experts. Regarding the standard deviation, since the mean plus and
minus three standard deviations will cover almost all observations, PErj - OPrj is
approximately equal to six standard deviations, which implies that brjσ= (PErj -
OPrj)/6. Every time a value is specified for αj, the probability that the constraint is
allowed to be violated, Model (7.11) is able to calculate an expected efficiency score
for a DMU. As αj increases, Φ-1(1 - αj) decreases, indicating that the feasible
region becomes larger. Consequently, the objective value becomes larger. In other
words, larger probabilities of allowing the constraints to be violated lead to higher
expected efficiency scores.
For more general cases in which the inputs Xij are also random variables
s m
expressed as xij, a similar transformation of zj = r = 1 ur yrj - u0 - i = 1 vi xij -
s m
r = 1 ur μrj - u0 - i = 1 vi μij = V j can be carried out, where μXij is the mean of
X
s m
xij and V j = var r = 1 ur yri - u0 - i = 1 vi xij : To transform the nonlinear form of
7.5 Interval Data 151
the standard deviation V j into a linear form in this case requires more restrictive
assumptions on the distributions of yrj and xij.
The assumption yrj = μrj + brjς makes V j a linear term. However, this
assumption implies that the correlation coefficient of any two random variables ypj
and yqj is:
σ pq bpj bqj σ 2
ρ ypj , yqj = = = 1,
σpσq bpj σ bqj σ
which is too restrictive to be satisfied in reality. It is thus better to use the nonlinear
term (7.10) for V j to calculate efficiency.
7.5 Interval Data
In real-world applications of DEA sometimes the data cannot be measured precisely

and is only known to lie in an interval. For example, the stocking level of a forest
cannot be measured precisely. There are also cases in which the data is missing or the
situation has not happened yet, and thus the data needs to be estimated. In these
cases, the measured efficiency will also be imprecise.
When data has an interval of values it is often called interval data. However, it
should be noted that the term interval data has been used in statistics to denote a
specific type of data that only allows for the degree of difference between items, but
not the ratio between them, e.g. the ratio of two temperatures or two dates is
meaningless. Since the term interval data is commonly used in the DEA literature
to denote data that covers a range of values, we will also use it, rather than coin a
new term.
_
Denote the interval data by “_” on top of the variable. Thus, we have X ij 2
_
X Lij , X U
ij and Y rj 2 Y Lrj , Y U
rj , where “L” and “U” denote the lower and upper
bounds of the interval, respectively. Every value in the interval can occur. Concep-
tually, the corresponding BCC input model can be formulated as:
_
s
r = 1 ur Y r0 - u0
_
E 0 = max: m _
i = 1 vi X i0
_
s
r = 1 ur Y rj - u0 ð7:12Þ
s:t: m _ ≤ 1, j = 1, . . . , n
i = 1 vi X ij
ur , vi ≥ ε, r = 1, . . . , s, i = 1, . . . , m
_ _ _
When X ij and Y rj have interval values, the calculated efficiency score E 0 will also
_
have interval value and is expressed as E 0 2 E L0 , EU 0 : Since {(X, Y) j X2 [X , X ],
L U
_
Y 2 [YL, YU]} is a convex and continuous set, E 0 , calculated from Model (7.10), will
0 : Based on this property, Kao (2006) proposed
be continuous in the range of E L0 , E U
the following two-level mathematical programming models to find the lower bound
E L0 and upper bound EU 0:
s
r = 1 ur yr0 - u0
E0 = max: m
i = 1 vi xi0
s
r = 1 ur yrj - u0
EL0 = min s:t: m ≤ 1, j = 1, . . . , n ð7:13aÞ
X Lij ≤ xij ≤ X U
ij i = 1 vi xij
Y Lrj ≤ yrj ≤ Y U
rj
ur , vi ≥ ε, r = 1, . . . , s, i = 1, . . . , m
8i, r, j u0 unrestricted in sign
s
r = 1 ur yr0 - u0
E 0 = max: m
i = 1 vi xi0
s
r = 1 ur yrj - u0
0 =
EU max s:t: m ≤ 1, j = 1, . . . , n ð7:13bÞ
X Lij ≤ xij ≤ X U
ij i = 1 vi xij
Y Lrj ≤ yrj ≤ Y U
rj
ur , vi ≥ ε, r = 1, . . . , s, i = 1, . . . , m
8i, r, j u0 unrestricted in sign
Given a set of xij and yrj values, the second level program calculates the
corresponding efficiency score, and the first level program determines the set of xij
and yrj values that will yield the smallest efficiency E L0 in Model (7.13a), and the
largest efficiency EU0 in Model (7.13b).
Models (7.13a) and (7.13b) are two-level programs, which cannot be solved
directly, and thus they must be transformed into one-level programs. Intuitively,
although this can also be proved theoretically, the lower bound E L0 occurs at the least
favorable conditions for the DMU being evaluated in measuring efficiency. That is,
the inputs and outputs occur at the upper and lower bounds, respectively, for this
DMU, and at the lower and upper bounds, respectively, for other DMUs. In contrast,
the upper bound EU 0 occurs at the most favorable conditions when measuring
efficiency. That is, the inputs and outputs occur at the lower and upper bounds,
respectively, for this DMU, and at the upper and lower bounds, respectively, for
other DMUs. Based on this reasoning, the two-level programs in Models (7.13a) and
(7.13b) are transformed into the following one-level programs:
7.6 Fuzzy Data 153
s
r = 1 ur Y r0 - u0
L
E L0 = max: m U
i = 1 vi X i0
s
r = 1 ur Y r0 - u0
L
s:t: m U ≤1
i = 1 vi X i0
s ð7:14aÞ
r = 1 ur Y rj - u0
U
m L ≤ 1, j = 1, . . . , n, j ≠ 0
i = 1 vi X ij
ur , vi ≥ ε, r = 1, . . . , s, i = 1, . . . , m
s
r = 1 ur Y r0 - u0
U
0 = max:
EU m L
i = 1 vi X i0
s
r = 1 ur Y r0 - u0
U
s:t: m L ≤1
i = 1 vi X i0
s ð7:14bÞ
r = 1 ur Y rj - u0
L
m U ≤ 1, j = 1, . . . , n, j ≠ 0
i = 1 vi X ij
ur , vi ≥ ε, r = 1, . . . , s, i = 1, . . . , m
This pair of models is of the conventional DEA type, which can be solved easily.
It should be noted that Models (7.13a) and (7.13b) have multiple solutions, in that
values of xij and yrj different from the lower and upper bounds may also produce EL0
0 : It_is thus _not appropriate to say that E 0 and E 0 always occur at the bound
and E U L U
values of X ij and Y rj .
7.6 Fuzzy Data
While using intervals to represent imprecise data is an intuitive approach, this

ignores the fact that some values in the interval are more likely to occur, while
others are less likely. If this phenomenon is to be considered, and there is no
knowledge of the probability distributions, then fuzzy numbers can be used to
express this type of data.
A fuzzy number X is a set, with a corresponding membership function μ ðxÞ
X
indicating the grade, in the range of zero and one, that the value x belongs to. A
typical fuzzy number is a trapezoidal one, expressed in short as (a, b, c, d ). The
corresponding membership function is:
Fig. 7.4 Frontiers for the Y

fuzzy-data example
4 B
3
C
2
B
1 A
O 1 2 3 4 5
ðx - aÞ=ðb - aÞ, a≤x≤b

1, b≤x≤c
μX~ ðxÞ =
ðd - xÞ=ðd - cÞ, c≤x≤d
0, otherwise
which indicates that the domain of this fuzzy number is (a, d ), in that a value smaller
than a or greater than d does not belong to this set, and values in the range of [b, c]
definitely belong to it. Larger grades of μ ðxÞ imply higher possibilities that the
X
value x belongs to this set. Conventionally “~” on top of a variable is used to denote a
fuzzy number.
Consider a very simple case of three DMUs, A, B, and C, depicted in Fig. 7.4,
where 1, 4, and 5 units of input X are used to produce 1, (2, 2.5, 3, 4), and 3 units of
output Y, respectively. The output of DMU B, Y B = (2, 2.5, 3, 4), is a trapezoidal
fuzzy number, with the membership function of:
ðy - 2Þ=0:5, 2 ≤ y ≤ 2:5
1, 2:5 ≤ y ≤ 3
μY~ B ðyÞ = ð7:15Þ
4 - y, 3≤y≤4
0, otherwise
For the output of DMU B, y, in the ranges of (2, 2.5), (2.5, 3), and (3, 4), the
frontiers are AC, ABC, and ABB’, respectively. These frontiers indicate that DMU
A is always efficient, DMU B is efficient in the second and third cases, and DMU C is
efficient in the first and second cases. To find the membership function for the fuzzy
efficiency of DMU B, EB , one only needs to find the membership grade μ ðeÞ
EB
7.6 Fuzzy Data 155
corresponding to y in the range of (2, 2.5). The efficiency of DMU B for y in the
range of (2, 2.5) is e = y/2.5, which implies y = 2.5e. The membership grade μ ðeÞ,
EB
according to Eq. (7.15), is thus (2.5e - 2)/0.5 = 5e - 4. For y greater than 2.5, the
efficiency of DMU B is e = 1, and the corresponding membership grade is the largest
of μ ðyÞ for y in this range, which, according to Eq. (7.15) is 1.
YB
To find the membership function for the fuzzy efficiency of DMU C, one only
needs to investigate the range of (3, 4) for the output of DMU B, y, because, based on
Eq. (7.15), μ ðeÞ = 1 for y in the range of (2, 3). For y in the range of (3, 4), the
EC
frontier is ABB’, and the efficiency of DMU C is e = 3/y. The membership grade of
y in this range is μ ðyÞ = 4 - y: Since e = 3/y implies y = 3/e, a result of
YB
μ ðeÞ = 4 - 3=e is obtained. To summarize, for Y B = (2, 2.5, 3, 4), DMU A is
EC
always efficient, with a crisp efficiency of EA = 1, and DMUs B and C have fuzzy
efficiencies, with the membership functions of:
5e - 4, 0:8 ≤ e ≤ 1 4 - 3=e, 3=4 ≤ e ≤ 1

μ ðeÞ = , μ ðeÞ =
EB 0, otherwise EC 0, otherwise
For this simple example the fuzzy efficiencies of all DMUs can be obtained
analytically. However, for general problems they will be too complicated to be
obtained analytically, and a numerical method must be relied on.
Since crisp values can be represented as degenerated fuzzy numbers with only
one value in their domain, we will assume all observations to be fuzzy for ease of
expression. Conceptually, the BCC input model with fuzzy observations can be
formulated as:
s ~
r = 1 ur Y r0 - u0
~ 0 = max:
E m ~
i = 1 vi X i0
s ~
r = 1 ur Y rj - u0 ð7:16Þ
s:t: ≤ 1, j = 1, . . . , n
m ~
i = 1 vi X ij
ur , vi ≥ ε, r = 1, . . . , s, i = 1, . . . , m
u0 unrestricted insign:
Other DEA models can be formulated similarly. Since the observations X ij and
Y rj are fuzzy numbers, the resulting efficiency E 0 is also a fuzzy number. The
fundamental fuzzy mathematical operations (Zimmermann 1996) are not able to
calculate E 0 from Model (7.16) directly, and special manipulations are needed.
When a fuzzy number, e.g. E 0 , is a result of the mathematical operations of some
other fuzzy numbers, e.g. X ij and Y rj , the membership function of E 0 can be obtained
from those of X ij and Y rj via the extension principle (Zadeh 1978; Yager 1986) in the
following form:
μ ðeÞ = supx,y min μ xij , μ yrj , 8 i, r, jje = E 0 ðx, yÞ ð7:17Þ

E0 X ij Y rj
where E0(x, y) is the conventional DEA model for measuring efficiency using
observations of (x, y). Based on (7.17), Kao and Liu (2000) developed a two-level
mathematical programming model to find the fuzzy efficiency measure E 0 .
The meaning of Expression (7.17) is that given a set of the (x, y) observations that
is able to yield the efficiency score e, the minimum of the membership grades
μ xij and μ yrj is determined, and μ ðeÞ is the largest of the minimum
X ij Y rj E0
membership grades corresponding to all the sets of (x, y) observations that are able
to produce the efficiency score e. Suppose μ ðeÞ = α: For those sets of (x, y)
E0
observations that have an efficiency score e, one must have μ xij ≥ α,
X ij
μ yrj ≥ α, and at least one xij or yrj with μ xij = α or μ yrj = α, to be able
Y rj X ij Y rj
L U L U
to result in μ ðeÞ = α: Let [ X ij α , X ij α ], [ Y rj α , Y rj α ], and [ðE0 ÞLα , ðE 0 ÞU
α ]
E0
denote the α-cuts of X ij , Y rj , and E 0 , respectively. Since all values in the α-cut have a
membership grade greater than or equal to α, ðE 0 ÞLα and ðE0 ÞU α are the smallest and
largest efficiency values, respectively, that have a membership grade α calculated
from the values of xij and yrj in their respective α-cuts, with at least one occurring at
the lower or upper bound of the α-cut, to have a membership grade α. ðE 0 ÞLα and
ðE 0 ÞU
α can thus be calculated as:
ðE 0 ÞLα = min:E0 ðx, yÞ

s:t: ðX ij ÞLα ≤ xij ≤ ðX ij ÞU
α, 8i, j
ðY rj ÞLα ≤ yrj ≤ ðY rj ÞU
α, 8r, j
ðE 0 ÞU
α = max:E 0 ðx, yÞ
s:t: ðX ij ÞLα ≤ xij ≤ ðX ij ÞU
α, 8i, j
α, 8r, j
which, in complete form, are:
s
r = 1 ur yr0 - u0
E0 = max: m
i = 1 vi xi0
s
r = 1 ur yrj - u0
ðE0 ÞLα = min s:t: m ≤ 1, j = 1, . . . , n ð7:18aÞ
ðX ij ÞLα ≤ xij ≤ ðX ij ÞU
α i = 1 vi xij
α
ur , vi ≥ ε, r = 1, . . . , s, i = 1, . . . , m
8i, r, j u0 unrestricted insign
7.6 Fuzzy Data 157
s
r = 1 ur yr0 - u0
E 0 = max: m
i = 1 vi xi0
s
r = 1 ur yrj - u0
ðE 0 ÞU
α = max s:t: m ≤ 1, j = 1, . . . , n ð7:18bÞ
ðX ij ÞLα ≤ xij ≤ ðX ij ÞU
α i = 1 vi xij
α
ur , vi ≥ ε, r = 1, . . . , s, i = 1, . . . , m
8i, r, j u0 unrestricted insign
Both of the mathematical programs are two-level ones, which must be

transformed into one-level programs before they can be solved. Similar to the case
of the interval data, in which the minimum and maximum efficiencies occur at the
least and most favorable conditions, respectively, the former occurs at the largest
inputs and smallest outputs of the DMU being evaluated, and the smallest inputs and
largest outputs of the other DMUs, and the latter occurs at the smallest inputs and
largest outputs of the DMU being evaluated, and the largest inputs and smallest
outputs of the other DMUs. The two-level mathematical programs of Models (7.18a)
and (7.18b) are thus transformed into the following one-level ones:
s L
r = 1 ur ðY r0 Þα - u0
ðE0 ÞLα = max: m U
i = 1 vi ðX i0 Þα
s L
s:t: m U ≤1
i = 1 vi ðX i0 Þα ð7:19aÞ
s U
r = 1 ur ðY rj Þα - u0
m L ≤ 1, j = 1, . . . , n, j ≠ 0
i = 1 vi ðX ij Þα
ur , vi ≥ ε, r = 1, . . . , s, i = 1, . . . , m
s U
ðE0 ÞU
α = max: m L
i = 1 vi ðX i0 Þα
s U
s:t: m L ≤1
i = 1 vi ðX i0 Þα ð7:19bÞ
s L
r = 1 ur ðY rj Þα - u0
m U ≤ 1, j = 1, . . . , n, j ≠ 0
i = 1 vi ðX ij Þα
ur , vi ≥ ε, r = 1, . . . , s, i = 1, . . . , m
Since all xij and yrj occur at either the lower or upper bounds of the α-cuts, the
condition of at least one xij or yrj occurring at either the lower or upper bound
required by the extension principle is satisfied. By enumerating various values of α,
different α-cuts of E0 are obtained, from which the membership function μ ðeÞ is
E0
constructed.
Pastor and Ruiz (2007) carried out a survey of the literature regarding negative data
in DEA models. Their work examined the slacks-based measure of Sharp et al.
(2007), semi-oriented radial measure of Emrouznejad et al. (2010), distance function
measure of Kerstens and van de Woestyne (2011), directional distance measure of
Branda (2015) and Diabat et al. (2015), range-adjusted measure of Tsang et al.
(2014), how Hatami-Marbini et al. (2014) handle negative data in intervals, and the
comments of Kerstens and van de Woestyne (2014) on Cheng et al.’s
method (2013).
The first study of ordinal data is probably Cook et al. (1993). However, this topic
was brought to the attention of a greater number of scholars after the work of Cooper
et al. (1999). Other studies on this topic include Despotis and Smirlis (2002), Zhu
(2003a, 2003b, 2004), Cook and Zhu (2006), Saen (2006), and Park (2007).
Qualitative data is relatively less studied in the literature and is sometimes discussed
with ordinal, interval, or fuzzy data. Huang and Chen (2013) proposed a different
approach for this type of data.
Based on the idea of chance-constrained programming presented in Charnes and
Cooper (1959, 1963), Land et al. (1994) and Olesen and Petersen (1995) developed
models to evaluate efficiency when the data is stochastic and can be represented by
random variables. Cooper et al. (2002) carried out a similar study. The chance-
constrained DEA has also been incorporated with birandom data (Tavana et al.
2014) and the satisficing concept of H. A. Simon to measure efficiency (Cooper et al.
1996). Keshvari and Kuosmanen (2013) discussed the connections between the
regression method of isotonic nonparametric least squares (INLS) and three different
types of DEA models and explored an idea for frontier estimation in stochastic
settings. When observations are random variables, the calculated efficiencies are
expected to be random variables as well. Kao and Liu (2009) applied a simulation
technique to find the distribution of the random efficiency when the observations are
independent, based on which the mean and variance are also calculated. Kao and Liu
(2019) repeated the same work by allowing the observations to be correlated.
Following the work of Cooper et al. (1999), interval data is usually discussed as a
type of imprecise data, as seen in the later studies of Jahanshahloo et al. (2009),
Khodabakhshi et al. (2010), Shokouhi et al. (2010), Inuiguchi and Mizoshita (2012),
Wei et al. (2014), and Azizi et al. (2015). Other studies include the idea of optimistic
and pessimistic measures of efficiency, as presented in Entani et al. (2002), and the
discriminant analysis proposed by Jahanshahloo et al. (2007).
Dozens of methods for measuring efficiency with fuzzy data have been proposed
in the literature, with Hatami-Marbini et al. (2011) providing a taxonomy and review
References 159
of most of these. However, new methods continued to be developed after this review,
with the work of Hougaard and Balezentis (2014) on the free disposal hull (FDH)
approach being an example of this.
References
Azizi, H., Kordrostami, S., & Amirteimoori, A. (2015). Slacks-based measures of efficiency in
imprecise data envelopment analysis: An approach based on data envelopment analysis with
double frontiers. Computers & Industrial Engineering, 79, 42–51.
Banker, R. D., & Morey, R. (1986). The use of categorical variables in data envelopment analysis.
Branda, M. (2015). Diversification-consistent data envelopment analysis based on directional-
distance measures. Omega, 52, 65–76.
Charnes, A., & Cooper, W. W. (1959). Chance-constrained programming. Management Science, 6,
73–79.
Charnes, A., & Cooper, W. W. (1963). Deterministic equivalences for optimizing and satisfying
under chance constraints. Management Science, 11, 18–39.
Cheng, G., Zervopoulos, P., & Qian, Z. (2013). A variant of radial measure capable of dealing with
negative inputs and outputs in data envelopment analysis. European Journal of Operational
Research, 225, 100–105.
Cook, W. D., Kress, M., & Seiford, L. M. (1993). On the use of ordinal data in data envelopment
analysis. Journal of the Operational Research Society, 44, 133–140.
Cook, W. D., & Zhu, J. (2006). Rank order data in DEA: A general framework. European Journal
of Operational Research, 174, 1021–1038.
Cooper, W. W., Deng, H., Huang, Z. M., & Li, S. X. (2002). Chance constrained programming
approaches to technical efficiencies and inefficiencies in stochastic data envelopment analysis.
Cooper, W. W., Huang, Z. M., & Li, S. X. (1996). Satisficing DEA models under chance
constraints. Annals of Operations Research, 66, 279–295.
Cooper, W. W., Park, K. S., & Yu, G. (1999). IDEA and AR-IDEA: Models for dealing with
imprecise data in DEA. Management Science, 45, 597–607.
Despotis, D. K., & Smirlis, Y. G. (2002). Data envelopment analysis with imprecise data. European
Diabat, A., Shetty, U., & Pakkala, T. P. M. (2015). Improved efficiency measure through directional
distance formulation of data envelopment analysis. Annals of Operations Research, 229,
325–346.
Emrouznejad, A., Anouze, A. L., & Thanassoulis, E. (2010). A semi-oriented radial measure for
measuring the efficiency of decision making units with negative data, using DEA. European
Entani, T., Maeda, Y., & Tanaka, H. (2002). Dual models of interval DEA and its extension to
interval data. European Journal of Operational Research, 136, 32–45.
Hatami-Marbini, A., Emrouznejad, A., & Agrell, P. J. (2014). Interval data with sign restrictions in
DEA. Applied Mathematical Modelling, 38, 2028–2036.
Hatami-Marbini, A., Emrouznejad, A., & Tavana, M. (2011). A taxonomy and review of the fuzzy
data envelopment analysis literature: Two decades in the making. European Journal of Oper-
ational Research, 214, 457–472.
Hougaard, J. L., & Balezentis, T. (2014). Fuzzy efficiency without convexity. Fuzzy Sets and
Systems, 255, 17–29.
Huang, W. T., & Chen, Y. W. (2013). Qualitative data envelopment analysis by affinity set: A
survey of subjective opinions for NPOs. Quality & Quantity, 47, 3079–3093.
Inuiguchi, M., & Mizoshita, F. (2012). Qualitative and quantitative data envelopment analysis with
interval data. Annals of Operations Research, 195, 189–220.
Jahanshahloo, G. R., Lotfi, F. H., Balf, F. R., & Rezai, H. Z. (2007). Discriminant analysis of
interval data using Monte Carlo method in assessment of overlap. Applied Mathematics and
Computation, 191, 521–532.
Jahanshahloo, G. R., Lotfi, F. H., Malkhalifeh, M. R., & Namin, M. A. (2009). A generalized model
for data envelopment analysis with interval data. Applied Mathematical Modelling, 33,
3237–3244.
Kao, C. (2006). Interval efficiency measures in data envelopment analysis with imprecise data.
European Journal Operational Research, 174, 1087–1099.
Kao, C., & Liu, S. T. (2009). Stochastic data envelopment analysis in measuring the efficiency of
Taiwan commercial banks. European Journal of Operational Research, 196, 312–322.
Kao, C., & Liu, S. T. (2019). Stochastic efficiency measures for production units with correlated
data. European Journal of Operational Research, 273, 278–287.
Kao, C. (2020). Measuring efficiency in a general production possibility set allowing for negative
data. European Journal of Operational Research, 282, 980–988.
Kao, C. (2022). Measuring efficiency in a general production possibility set allowing for
negative data: An extension and a focus on returns to scale. European Journal of Operational
Research, 296, 267–276.
Kao, C., & Lin, P. H. (2011). Qualitative factors in data envelopment analysis: A fuzzy number
Kao, C., & Liu, S. T. (2000). Fuzzy efficiency measures in data envelopment analysis. Fuzzy Sets
and Systems, 113, 427–437.
Keshvari, A., & Kuosmanen, T. (2013). Stochastic non-convex envelopment of data: Applying
isotonic regression to frontier estimation. European Journal of Operational Research, 231,
481–491.
Kerstens, K., & Van de Woestyne, I. (2011). Negative data in DEA: A simple proportional distance
function approach. Journal of the Operational Research Society, 62, 1413–1419.
Kerstens, K., & Van de Woestyne, I. (2014). A note on a variant of radial measure capable of
dealing with negative inputs and outputs in DEA. European Journal of Operational Research,
234, 341–342.
Khodabakhshi, M., Gholami, Y., & Kheirollahi, H. (2010). An additive model approach for
estimating returns to scale in imprecise data envelopment analysis. Applied Mathematical
Modelling, 34, 1247–1257.
Kumbhaker, S. C., & Lovell, C. A. K. (2000). Stochastic frontier analysis. Cambridge University
Press.
Land, K., Lovell, C. A. K., & Thore, S. (1994). Chance-constrained data envelopment analysis.
Managerial and Decision Economics, 14, 541–554.
Olesen, O. B., & Petersen, N. C. (1995). Chance constrained efficiency evaluation. Management
Science, 41, 442–457.
Park, K. S. (2007). Efficiency bounds and efficiency classifications in DEA with imprecise data.
Pastor, J., & Ruiz, J. (2007). Variables with negative values in DEA. In J. Zhu & W. D. Cook (Eds.),
Modeling data irregularities and structural complexities in data envelopment analysis
(pp. 63–84). Springer.
Portela, M. C. A. S., Thanassoulis, E., & Simpson, G. (2004). Negative data in DEA: A directional
distance approach applied to bank branches. Journal of the Operational Research Society, 55,
1111–1121.
Saen, R. F. (2006). Technologies ranking in the presence of cardinal and ordinal data. Applied
Mathematics and Computation, 176, 476–487.
Sharp, J. A., Meng, W., & Liu, W. (2007). A modified slacks-based measure model for data
envelopment analysis with ‘natural’ negative outputs and inputs. Journal of the Operational
References 161
Shokouhi, A. H., Hatami-Marbini, A., Tavana, M., & Saati, S. (2010). A robust optimization
approach for imprecise data envelopment analysis. Computers & Industrial Engineering, 59,
387–397.
Tavana, M., Shiraz, R. K., & Hatami-Marbini, A. (2014). A new chance-constrained DEA model
with birandom input and output data. Journal of the Operational Research Society, 65,
1824–1839.
Tsang, S. S., Chen, Y. F., Lu, Y. H., & Chiu, C. R. (2014). Assessing productivity in the presence of
negative data and undesirable outputs. Service Industries Journal, 34, 162–174.
Wei, Q., Chang, T. S., & Han, S. (2014). Quantile-DEA classifiers with interval data. Annals of
Operations Research, 217, 535–563.
Yager, A. (1986). A characterization of the extension principle. Fuzzy Sets and Systems, 18,
205–217.
Zadeh, L. A. (1978). Fuzzy sets as a basis for a theory of possibility. Fuzzy Sets and Systems, 1,
3–28.
Zhu, J. (2003a). Imprecise data envelopment analysis (IDEA): A review and improvement with an
application. European Journal of Operational Research, 144, 513–529.
Zhu, J. (2003b). Efficiency evaluation with strong ordinal input and output measures. European
Zhu, J. (2004). Imprecise DEA via standard linear DEA models with a revisit to a Korean mobile
telecommunication company. Operations Research, 52, 323–329.
Zimmermann, H. J. (1996). Fuzzy set theory and its applications (3rd. ed.). Kluwer-Nijhoff.
Chapter 8
Changes of Efficiency over Time
The DEA technique measures efficiency in a relative manner, in that the perfor-
mances of the DMUs in a group are compared with each other. The efficient ones
may not be efficient when compared with the DMUs of other groups. Similarly, the
inefficient ones may become efficient when compared with those of other groups.
The efficiency measures for DMUs of different groups are thus not comparable.
The major objective of efficiency measurement is to identify the inefficient
DMUs and make improvements so that they can perform better in the future.
However, if every inefficient DMU improves their performance, they may still be
evaluated as inefficient, with even lower efficiency scores, because the measurement
is relative, rather than absolute. A method for measuring the improvement or decline
of a DMU from one period to another is thus desired. This is especially important for
examining the effect of an act or policy over a period of time, such as the reorga-
nization of Taiwanese forests in 1989 (Kao 2000) and the Sarbanes-Oxley Act of the
USA in 2002 (Chang et al. 2009). In this chapter we will introduce two methods to
achieve this: the Malmquist productivity index (MPI) and Luenberger productivity
index (LPI). The MPI is based on the radial distance function, which has the input-
and output-orientations. The LPI is based on the directional distance function, in that
both the inputs and outputs are considered. Both of the indexes can be measured
under constant and variable returns to scale, and different versions have appeared in
the literature.
In the next section we will discuss the theoretic foundations of the MPI. Then
how the MPI and an important version of it, the global MPI, are calculated using the
DEA technique will be described. After that, the LPI is introduced.
https://doi.org/10.1007/978-3-031-27593-7_8
164 8 Changes of Efficiency over Time
8.1 Theoretic Foundation of MPI
The Malmquist productivity index (MPI) was introduced by Caves et al. (1982) to
compare the productivity of two DMUs, where the two DMUs may represent the
same DMU at two different times, or two different DMUs at the same time or
different times, based on the idea proposed in Malmquist (1953) of constructing
quantity indexes as ratios of distance functions. In addition to a productivity index,
input and output indexes were also defined. Each DMU is permitted to have its own
production function and is operating on their production frontiers, i.e. they are
efficient.
8.1.1 Input Index
Let T p be the production technology of DMU p that transforms inputs X p into

outputs Y p, i.e., T p = {(X p, Y p) j Y p ≥ 0 can be produced by X p ≥ 0}. Recall that
the input distance function based on the technology of DMU p is
1
DpI ðY, X Þ = ð8:1Þ
min:fθjðθX, Y Þ 2 T p g
Note that the notation in this chapter is slightly different from that of Chap. 3 to be
consistent with that used in Caves et al. (1982). The DMU p Malmquist input index
is defined as:
DpI ðY p , Xp Þ
M pI ðX p , Xq Þ = ð8:2Þ
DpI ðY p , Xq Þ
which measures the relative difference between the input vectors of DMUs p and q,
given that the output vector is that of DMU p. Since DpI ðY p , X p Þ = 1 (because DMU
p is efficient), M pI ðX p , Xq Þ becomes min. {θ| (θXq, Y p) 2 T p}, which is the minimum
factor θ required to expand the input vector of DMU q onto the production surface of
DMU p, given that the output vector is that of DMU p. If M pI ðX p , Xq Þ > 1, then the
input vector of DMU p is larger than that of DMU q, from the perspective of DMU
p’s technology.
The DMU q Malmquist input index can be defined similarly and derived as
M qI ðXp , X q Þ = DqI ðY q , Xp Þ=DqI ðY q , Xq Þ = DqI ðY q , Xp Þ = ½min:fθjðθX p , Y q Þ 2 T q g - 1 :
It is therefore the maximum θ required to contract the input vector of DMU p onto the
production surface of DMU q, given that the output vector is that of DMU q. If it is
greater than one, then this implies that the input vector of DMU p is larger than that of
DMU q, from the perspective of DMU q’s technology. Caves et al. (1982) showed
that the geometric mean of the Malmquist input indexes evaluated at DMUs p and q is
8.1 Theoretic Foundation of MPI 165
equal to the Törnqvist index for comparing the inputs of DMUs p and q, provided
they have translog distance functions.
8.1.2 Output Index
From the output side, the distance function based on the technology of DMU p is
1
DpO ðY, X Þ = ð8:3Þ
max:fϕjðX, ϕY Þ 2 T p g
Caves et al. (1982) defined the DMU p Malmquist output index as:
DpO ðY p , X p Þ
M pO ðY p , Y q Þ = ð8:4Þ
DpO ðY q , X p Þ
to measure the relative difference between the output vectors of DMUs p and q,
given that the input vector is that of DMU p. Based on the definition of M pO , this
expression can be simplified to max. {ϕ| (X p, ϕYq) 2 T p}, which states that the DMU
p Malmquist output index is the maximum factor φ required to expand the output
vector of DMU q onto the production surface of DMU p, given that the input vector
is that of DMU p. A value of greater than one indicates that the output vector of
DMU p is larger than that of DMU q from the viewpoint of DMU p.
Similarly, the DMU q Malmquist output index is defined as M qO ðY p , Y q Þ =
DO ðY p , Xq Þ=DqO ðY q , X q Þ, which is equal to DqO ðY p , X q Þ, or [ max . {ϕ j (Xq, ϕ
q
Y p)2 Tq}]-1, as DqO ðY q , X q Þ = 1: M qO ðY p , Y q Þ is thus the minimum factor φ required

to contract the output vector of DMU p onto the production surface of DMU q, given
that the input vector is that of DMU q. If this index is greater than one, then the
output vector of DMU p is larger than that of DMU q from the perspective of DMU
q. Caves et al. (1982) also proved that if DMUs p and q have translog output distance
functions, then the geometric mean of the Malmquist output indexes evaluated at
DMUs p and q is equal to the Törnqvist index for comparing the outputs of DMUs
p and q.
8.1.3 Productivity Index
In addition to the input and output indexes for comparing the inputs and outputs,
respectively, of two DMUs, Caves et al. (1982) also defined the productivity index to
measure productivity differences between two DMUs. There are two aspects for this
measurement: input and output.
The DMU p Malmquist input-based productivity index is given by:
DpI ðY q , Xq Þ
MPIpI ðXp , Y p , Xq , Y q Þ = ð8:5Þ
DpI ðY p , Xp Þ
which is equal to DpI ðY q , X q Þ due to DpI ðY p , X p Þ = 1. Since DpI ðY q , X q Þ = ½ min:fθ j

(θXq, Yq) 2 T p}]-1, MPI pI is the maximum input contraction factor, such that the
contracted input for DMU q and the DMU q output vector lie on the production
surface of DMU p. If this index is greater than one, then DMU p has a higher
productivity level than DMU q. Similarly, the DMU q Malmquist input-based
productivity index is MPIqI ðXp , Y p , Xq , Y q Þ = DqI ð Y q , X q Þ=DqI ð Y p , Xp Þ =
q
1=DI ðY , X Þ = min:fθjðθX , Y p Þ 2 T q g, which is the minimum input expansion
p p p
factor, such that the expanded input for DMU p and the DMU p output vector lie on
the production surface of DMU q. If this is greater than one, then DMU p has a
higher productivity level than DMU q. Caves et al. (1982) proved that the geometric
mean of the DMU p and q Malmquist input-based productivity index is equal to the
Törnqvist productivity index corrected by a scale factor, provided DMUs p and
q have translog input distance functions.
The Malmquist input-based productivity index is similar to the Malmquist input
index. By comparing the definition of these two indexes, Expressions (8.2) and (8.5),
it is noted that the input index measures the increase in the inputs of a DMU while
fixing the outputs at the same level. If this index is greater than one, then a larger
amount of the inputs of this DMU have been consumed to produce the same level of
outputs, which is less efficient. The input-based productivity index measures the
reduction in the inputs of a DMU used in producing the current level of outputs. A
value of greater than one implies that this DMU is more productive.
The Malmquist productivity index can also be defined from the output side as:
DpO ðY p , Xp Þ
MPIpO ðXp , Y p , Xq , Y q Þ = ð8:6Þ
DpO ðY q , Xq Þ
which is equal to 1=DpO ðY q , X q Þ, due to DpO ðY p , X p Þ = 1: Based on the definition of

DpO ðY q , Xq Þ = ½max:fϕjðXq , ϕY q Þ 2 T p g - 1 , MPIpO is the maximum expansion
factor for the DMU q output vector, such that the resulting expanded output vector
and the DMU q input vector are just on the production surface of DMU p. DMU
p thus has a higher level of productivity than DMU q from the perspective of DMU
p’s production structure if MPIpO > 1. If the base DMU is changed from p to q, then
the productivity index becomes MPIqO ðX p , Y p , X q , Y q Þ = DqO ð Y p , X p Þ=
DO ð Y , X Þ = DO ðY , X Þ = ½max:fϕjðX , ϕY p Þ 2 T q g - 1 , which is the minimum
q q q q p p p
output contraction factor, such that the contracted output vector for DMU p and the
DMU p input vector are just on the production surface of DMU q. If this index is
greater than one, then DMU p has a higher level of productivity than DMU q from
the perspective of DMU q’s production structure. Similar to the input-based case, the
geometric mean of the DMU p and DMU q Malmquist output-based productivity
indexes is equal to the Törnqvist productivity index corrected by a scale factor if
DMUs p and q have translog output distance functions. The difference between the
8.2 DEA Measurement of MPI 167
Malmquist output index and the Malmquist output-based productivity index can be
explained similar to that of the input case.
8.2 DEA Measurement of MPI
Caves et al. (1982) assumed the DMUs to be efficient, and the form of the production
functions to be known, although the parameters of the functions needed not be given
in deriving the relationships between the Malmquist and Törnqvist indices. Färe
et al. (1994) applied the DEA technique to calculate the Malmquist productivity
index for the same DMU at two different periods. The same set of DMUs with
observations at two periods was used to construct the empirical production frontier at
the two periods. The DMUs were not required to be efficient. Since the nonpara-
metric DEA technique was used, there was no need to specify a function form for the
technology.
The Malmquist productivity index proposed by Färe et al. (1994) can be
discussed from either the input or output sides. Recall that the output-based produc-
tivity index defined for a DMU at periods t and t + 1 by Caves et al. (1982) is
MPItO X t , Y t , X tþ1 , Y tþ1 = DtO ðY t , Xt Þ=DtO Y tþ1 , X tþ1 : Compared with Expres-
sion (8.6), it is noted that DMUs p and q have been replaced with periods t and
t + 1, respectively. If MPItO > 1, then the DMU being investigated has a higher
productivity at period t than at period t + 1, based on the technology of period t. Färe
et al. (1994) exchanged the periods of comparison, such that the Malmquist produc-
tivity index is greater than one when the productivity of this DMU has increased
from period t to period t + 1. Therefore, the Malmquist output-based productivity
index defined by Färe et al. (1994) becomes:
DtO Y tþ1 , X tþ1

MPItO Xt , Y t , X tþ1 , Y tþ1 = ð8:7Þ
DtO ðY t , X t Þ
The distance function of DtO ðY t , X t Þ, by the definition of (8.3), is the efficiency of

a DMU at period t based on the technology of period t, whose value is less than or
equal to one. The distance function of DtO Y tþ1 , X tþ1 =
- 1
max: ϕj X tþ1 , ϕY tþ1 2 T t is the efficiency of a DMU at period t + 1
measured from the technology of period t. Under the assumption of variable returns
-1
to scale, DtO Y tþ1 , X tþ1 can be calculated via the following linear program:
max:φ
n
s:t: Σ λj X itj ≤ X tþ1
i0 , i = 1, . . . , m
j=1
n
Σ λj Y rt j ≥ φY tþ1
r0 , r = 1, . . . , s
j=1 ð8:8Þ
n
Σ λj = 1
j=1
λj ≥ 0, j = 1, . . . , n
It is possible that the efficiency index φ is greater than one, as long as the
observation of this DMU at period t + 1 lies outside of the production frontier of
period t. The MPItO , which is the ratio of DtO Y tþ1 , X tþ1 to DtO ðY t , X t Þ, with a value
greater than one thus implies that the efficiency of this DMU has increased from
period t to period t + 1. Conversely, if it is less than one, then the efficiency has
declined. The efficiency remains the same if it is equal to one.
As discussed in Caves et al. (1982), the Malmquist productivity index can also be
measured based on the technology of period t + 1, and the formulation of Färe et al.
(1994) is:
Dtþ1 Y tþ1 , Xtþ1

MPItþ1
O Xt , Y t , X tþ1 , Y tþ1 = O
ð8:9Þ
O ðY , X Þ
Dtþ1 t t
The numerator is the efficiency of a DMU at period t + 1 based on the technology

of period t + 1, and the denominator is the efficiency index of this DMU at period
t measured from the technology of period t + 1. The corresponding linear program is:
1
= max:φ
O ðY , X Þ
Dtþ1 t t
n
s:t: λj X tþ1
i j ≤ X i0 ,
t
i = 1, . . . , m
j=1
n
λj Y tþ1
r j ≥ φY r0 ,
t
r = 1, . . . , s ð8:10Þ
j=1
n
λj = 1
j=1
λj ≥ 0, j = 1, . . . , n
The value of φ can be greater than, equal to, or less than one.
To see the difference between MPItO and MPItþ1 O , consider DMU A with the
observations of At and At + 1, and the production frontiers of the two periods, as
8.2 DEA Measurement of MPI 169
Fig. 8.1 Output-based MPI Y
Aˆ t +1 Period t+1
t
Â
A t +1 Period t
A t +1
At
At
O D B C X
depicted in Fig. 8.1. The projection points of At and At + 1 on the frontiers of periods
t tþ1 t tþ1
t and t + 1 are A , A and A , A , respectively. The MPIs based on the technology
tþ1 t
of periods t and t + 1 are MPItO At , Atþ1 = Atþ1 C=A C = At B=A B and
tþ1 t
MPItþ1
O At , Atþ1 = Atþ1 C=A C = At B=A B , respectively. In the followings
the notation of Xt, Yt, Xt + 1, and Yt + 1 in MPItþ1
O X t , Y t , X tþ1 , Y tþ1 is simplified
to Z and Z
t t+1
, as MPIO Z , Z
tþ1 t tþ1
. While these two values may not be the same,
they usually have the same trend of being greater or less than one to provide
consistent conclusions with regard to productivity changes.
However, there are still cases in which different technologies yield inconsistent
results. Consider the four DMUs depicted in Fig. 8.2, with the data shown in
Table 8.1. There are two points to be noted. One is that although DMUs A, B, and
D are efficient in both periods based on their respective frontiers, their productivities
have changed when the same technology is used for measurement. The other is that
the MPIs measured from different technologies may be inconsistent. The results can
be classified into three types. The first is that the MPIs measured from different
technologies have the same value, for example, DMU A, where the MPI is consis-
tently equal to two, regardless of the period used for measurement. The second is that
the MPIs measured from different technologies have different values, yet the trend is
the same. DMU D is an example of this type, where the MPIs are different (9/7
versus 26/21), yet they consistently indicate that the productivity has increased. The
third is that the MPIs measured from different technologies are different and also
show different trends. DMUs B and C are of this type, where the period t technology
indicates that both DMUs have an increased productivity, whereas the period t + 1
technology shows a decreased productivity for both DMUs.
In order to avoid choosing different benchmarks, and so obtaining different
results, Färe, Grosskopf, Norris, and Zhang (FGNZ) (1994) suggested using the
geometric mean of the two MPIs to be the final MPI:
Fig. 8.2 Measurement of Y

the MPI
10
D t +1
B t +1 Period t+1
8
C t +1
t
Period t
B Dt
6 C t
A t +1
4
2 At
O 2 4 6 8 10 12 X
1=2
DtO Ztþ1 Dtþ1 Ztþ1
MPIFGNZ Zt , Ztþ1 = × Otþ1 t ð8:11Þ
O
DO ðZ Þ
t t
DO ðZ Þ
This index not only better represents the productivity changes, but also can be
decomposed into two components to provide more information. The
decomposition is:
1=2
Dtþ1 Ztþ1 DtO Ztþ1 DtO ðZt Þ
MPIFGNZ Zt , Ztþ1 = O
×
O
DO ðZt Þ
t
Dtþ1
O Ztþ1 O ðZ Þ
Dtþ1 t
= ðefficiency changeÞ × ðtechnical changeÞ
The first part is the ratio of two efficiencies measured at periods t + 1 and t and is
called the efficiency change or catching up (to the frontier). The second part is the
geometric mean of two ratios, which captures the shift in technology between the
two periods from the viewpoint of the DMU at the two periods, and this is called the
technical change. Using DMU A in Fig. 8.1 to explain this, the efficiency change is
the ratio of the efficiency at period t + 1 to that at period t, which is
tþ1 t
Atþ1 C=A C = At B=A B : The technical changes evaluated at At + 1 and At are
tþ1 tþ1 tþ1 tþ1 t t
Atþ1 C=A C = Atþ1 C=A C = A C=A C and At B=A B = At B=A B
t t tþ1 tþ1
= A B=A BÞ, respectively, and whose geometric mean, A C=A C
t t 1=2
A B=A B , is an average technical change.
Measuring the efficiency of a DMU based on the frontier constructed from other
DMUs may be infeasible. For example, if the efficiency index of DMU E in period
t in Fig. 8.1 is to be measured based on the technology of period t + 1, then it will be
infeasible because the frontier of period t + 1 starts from D, and there is no frontier
8.2
DEA Measurement of MPI
Table 8.1 Data and the MPIs for an example

t
DMU (Xt, Yt) (Xt + 1, Yt + 1) DtO ðZ t Þ DtO Z tþ1 MPI tO Dtþ1
O ðZ Þ Dtþ1
O Z tþ1 MPI tþ1
O MPI FGNZ
O
A (2, 2) (2, 4) 1 2 2 1/2 1 2 2
B (4, 6) (8, 8) 1 6/5 6/5 9/8 1 8/9 16=15
C (6, 6) (9, 7) 18/19 42/41 133/123 9/10 21/25 14/15 1.0046
D (10, 7) (11, 9) 1 9/7 9/7 21/26 1 26/21 1.2617
171
facet of period t + 1 for which point E can be used to measure the (output) efficiency.
The associated linear program will always be feasible only if the technology is
constant (or non-increasing) returns to scale.
In addition to feasibility, there is another advantage to measuring the MPI using
the constant returns to scale technology (Färe et al. 1994), and this is the technical
and scale efficiency decomposition. As discussed in Chap. 2, the ratio of the CCR
efficiency, which is measured under the constant returns to scale technology, to the
BCC efficiency, which is measured under the variable returns to scale technology, is
the scale efficiency. Suppose the distance functions in (8.7) and (8.9) are measured
under constant returns to scale. The MPI in (8.11) can then be further
decomposed as:
1=2
TEtþ1 SE tþ1 DtO Ztþ1 DtO ðZt Þ
MPIFGNZ Z ,Z
t tþ1
= t × ×
O
TE SEt Dtþ1
O Ztþ1 O ðZ Þ
Dtþ1 t
= ðtechnical efficiency changeÞ × ðscale efficiency changeÞ × ðtechnical changeÞ
To measure the efficiency based on the constant returns to scale technology, one
n
simply deletes the convex combination constraint j = 1 λj = 1 from the
corresponding linear program.
The Malmquist productivity index can also be measured from the input side, as
Caves et al. (1982) did, which is MPItI Zt , Ztþ1 = DtI Ztþ1 =DtI ðZt Þ, with period t as
the base period. The input-based productivity of a DMU at period t is higher than that
at period t + 1 if MPItI > 1. To express the productivity change of a DMU from
period t to t + 1, the input-based MPI becomes:
DtI ðZt Þ
MPItI Zt , Ztþ1 = ð8:12Þ
DtI Ztþ1
A value greater than one indicates that the productivity of this DMU has increased
from period t to period t + 1. According to the definition of the input distance
function presented in Model (8.1), MPItI is actually the ratio of the input efficiency of
a DMU at period t + 1 to that at period t.
Referring to DMU A in Fig. 8.3, its input-based MPI is MPItI At , Atþ1 =
tþ1 t
DtI ðAt Þ=DtI Atþ1 = A C=Atþ1 C = A B=At B : The MPI based on the technol-
ogy of period t + 1 is
tþ1 t
tþ1 t
MPII A , Atþ1
= DI ðA Þ=DI A
tþ1 t tþ1 tþ1
= A C=A C = A B=At B .
tþ1
The
MPI defined by Färe et al. (1994) is the geometric mean of MPItI and MPItþ1 I ,
which, similar to the output case, can also be decomposed into the product of
efficiency change and technical change under the variable returns to scale technol-
ogy, and the efficiency change can be further decomposed into the product of the
technical efficiency change and scale efficiency change under constant returns to
scale technology.
8.3 Global Malmquist Productivity Index 173
Fig. 8.3 Input-based MPI Y
Period t+1
Period t
Aˆ t +1
C At +1
A t +1
Ât
B
At At
O X
8.3 Global Malmquist Productivity Index
The MPI of Färe et al. (1994) takes the geometric mean of two MPIs measured at two
periods as the final MPI to alleviate the inconsistency of the two MPIs. However, it
loses the property of circularity that is possessed by the MPI measured from one base
period. Suppose there are three periods, t - 1, t, and t + 1, between each pair the one-
base-period MPIs are measured. The output-based MPIs for the pairs of periods (t-
1, t), (t, t + 1), and (t-1, t + 1) based on the technology of period t-1 are:
DtO- 1 ðZt Þ
MPItO- 1 Zt - 1 , Zt =
DtO- 1 Zt - 1
DtO- 1 Ztþ1
MPItO- 1 Zt , Ztþ1 =
DtO- 1 ðZt Þ
DtO- 1 Ztþ1
MPItO- 1 Zt - 1 , Ztþ1 =
DtO- 1 Zt - 1
The product of the MPIs for the pairs of periods (t-1, t) and (t, t + 1) is equal to
that for (t-1, t + 1), i.e. MPItO- 1 Z t - 1 , Z t × MPIO t-1
Z t , Z tþ1 =
t-1 t-1
MPIO Z , Z tþ1
: This is the property of circularity, which certainly holds for
any three periods, based on the technology of any period.
In either the MPI of one base period or the average MPI of two base periods, the
distance measures can be infeasible under variable returns to scale. The MPIs
measured from two different base periods may be different or even inconsistent, in
that one indicates an improvement while the other indicates deterioration. The
average MPI of two base periods alleviates the inconsistency of the two one-base-
period MPIs, however, whether the average MPI is the correct one is still
questionable. To solve these problems, Pastor and Lovell (2005, 2007) proposed a
global MPI, which uses the observations of all periods to construct a global frontier,
for measurement.
Suppose there are T periods. The idea of the global MPI is to use the observations
Zt = (Xt, Yt), t = 1, . . ., T to construct the frontier. The output distance for a DMU at
period k under variable returns to scale in this case is:
1
= max:φ
DOG ðZk Þ
T n
s:t: λjt X itj ≤ X i0k , i = 1, . . . , m
t=1j=1
T n
λjt Y rt j ≥ φY r0
k
, r = 1, . . . , s ð8:13Þ
t=1j=1
T n
λjt = 1
t=1j=1
λjt ≥ 0, j = 1, . . . ,n, t = 1, . . . , T
O Z
k
Since all observations are used to construct the frontier, DG is always
feasible, with φ ≥ 1. The global MPI for a DMU between periods t and t + a is
defined as:
O ðZ Þ
tþa
DG
O ðZ , Z
MPIG Þ= ð8:14Þ
t tþa
DO ðZ Þ
G t
If it is greater than one, then the productivity of this DMU has increased from
period t to period t + a. Since there is only one frontier, and every DMU at any period
uses this frontier to measure efficiency, the problem of inconsistency does not exist.
Regarding circularity, as the product of the global MPIs of a DMU for the pairs of
periods (t, t + a) and (t + a, t + a + b) is equal to that for (t, t + a + b), the property of
circularity is satisfied:
O ðZ Þ
tþa
DG
O ðZ , Z Þ=
t tþa
MPIG
DO ðZ Þ
G t
O ðZ Þ
tþaþb
DG
O ðZ , Ztþaþb Þ =
tþa
MPIG
DO ðZ Þ
G tþa
O ðZ Þ
tþaþb
DG
O ðZ , Z Þ=
t tþaþb
MPIG
DO ðZ Þ
G t
8.3 Global Malmquist Productivity Index 175
Fig. 8.4 Global MPI Y
10
D t +1
Global
Cˆ t +1
8 B t +1
Ĉ t
Cˆ t –1 C t +1 Dt
Bt
6 Ct
t –1
D t –1
C
4 A t +1
B t –1
2
A t –1 At
E F G
O 2 4 6 8 10 12 X
Consider four DMUs with observations at three periods, t - 1, t, and t + 1 as

depicted in Fig. 8.4, where the observations at periods t and t + 1 are the same as
those in Fig. 8.2. The global frontier is the line segments At - 1At + 1BtBt + 1Dt + 1. We
will discuss the problems of feasibility, consistency, and circularity using the global
frontier.
If the output-based distance of DMU At - 1 is to be measured based on the
technology of either period t or period t + 1, it will be infeasible. In contrast, it is one
under the global technology, and in fact, all DMUs at any period have a feasible
distance measure based on the global frontier. For the inconsistent MPI measures of
DMUs B and C between periods t and t + 1, their MPIs based on the period
t technology, referring to Table 8.1, are 6/5 and 133/123, respectively, and based
on the period t + 1 technology are 8/9 and 14/15, respectively, which result in
geometric means of 16=15 and 1.0046 (or 1862=1845 ), respectively. The
averages indicate that the productivities of both DMUs B and C have increased,
although to a limited extent. Their global MPIs, however, are:
DG
O B
tþ1
1
MPIG t
O B ,B
tþ1
= = =1
D O ðB Þ
G t 1
tþ1
DG C tþ1 C tþ1 G=C G 21=25 49
MPIG t
C ,C tþ1
= OG t = = =
O
DO ðC Þ t
Ct F=C F 6=7 50
which indicate that the productivity of DMU B remains the same and that of DMU
C actually decreases. Both are different from the conclusions of the average MPI.
DMU C is used as an example to explain circularity. The global MPIs between
each pair of periods are:
t
t-1 C t F=C F
MPIG
O C , Ct = t-1
C t - 1 E=C E
tþ1 tþ1
C G=C G
MPIG t
O C ,C
tþ1
= t
t
C F=C F
tþ1
t-1 C tþ1 G=C G
MPIG
O C , C tþ1 = t-1
t-1
C E=C E
t-1
As expected, the product of MPIG O C , Ct and MPIG t
O C ,C
tþ1
is equal to
t-1
G
MPIO C , C tþ1
:.
Using one global frontier to measure the MPI avoids several undesirable situa-
tions that can occur when measuring these from several frontiers. However, the
single global frontier also makes the decomposition of efficiency and technical
changes impossible.
8.4 Luenberger Productivity Index
The Malmquist productivity index is based on the radial distance function of

Shephard (1970), and there are two forms, input- and output-based, where the former
seeks the smallest contracted inputs at which the current level of outputs can still be
produced, and the latter seeks the largest expanded outputs that can still be produced
with the current level of inputs. Either the inputs or outputs are fixed, and thus both
cannot be improved at the same time. In contrast, there is a directional distance
function, as discussed in Chap. 3, where both the inputs and outputs can be altered at
the same time, and an index is defined to measure the productivity changes between
two periods. Since the directional distance function is similar to the shortage
function of Luenberger (1992), the corresponding index is called the Luenberger
productivity index (LPI) by Chambers et al. (1996) and Chambers and Pope (1996).
Recall that the directional distance function with respect to direction d = (f, g), as
defined in Expression (3.18) of Chap. 3, is:
DðX, Y, f , gÞ = max:fηjðX - ηf , Y þ ηgÞ 2 T g: ð8:15Þ
A DMU is efficient if η = 0,Z and inefficient if η > 0, and a larger η* implies

that it is less efficient. To simplify the notation, let Z = (X, Y). The directional
distance for a DMU at period a moving along the direction da toward the period
→b
b frontier is D ðZa , da Þ, and is measured as:
8.4 Luenberger Productivity Index 177
max:η
n
s:t: λj X ibj ≤ X i0a - ηf ia , i = 1, . . . , m
j=1
n
λj Y rbj ≥ Y r0
a
þ ηgra , r = 1, . . . , s
j=1 ð8:16Þ
n
λj = 1
j=1
λj ≥ 0, j = 1, . . . , n
When using the period b technology to evaluate the inefficiency of a DMU at

period a, it is possible that η < 0. This indicates that this DMU lies outside of the
period b frontier and is considered as superefficient.
The LPI for a DMU between periods b and b + d based on the period
b technology is:
→b →b
Lb Zb , db , Zbþd , dbþd = D Zb , db - D Zbþd , dbþd ð8:17Þ
If it is positive, then the productivity of the DMU has increased from period b to
period b + d; if not, it has decreased. This index also possesses the property of
circularity, in additive form, as Lt(Zt, dt, Zt + b, dt + b) + Lt(Zt + b, dt + b, Zt + b + d,
dt + b + d)= Lt(Zt, dt, Zt + b + d, dt + b + d). Chambers et al. (1996) defined the LPI as the
arithmetic average of two LPIs based on the two periods b and a being compared.
That is:
1 b b b a a
LPI Zb , db , Za , da = L Z , d , Z , d þ La Zb , db , Za , da
2
→b →b →a →a
ð8:18Þ
1
= D Zb , db - D ðZa , da Þ þ D Zb , db - D ðZa , da Þ
2
Taking the average of two LPIs measured from the two periods will alleviate the
inconsistency between the two LPIs.
Consider DMU A in Fig. 8.5, faced with the technologies of periods t and t + 1,
where At and At + 1 are the observations of DMU A at periods t and t + 1, respectively.
Suppose the direction is set to dt = ðf t , gt Þ = ðX At , Y At Þ and dt + 1= f tþ1 , gtþ1 =
ðX Atþ1 , Y Atþ1 Þ: Then the directional distances for At and At + 1 based on the period
t technology are ηt and ηtþ1 , respectively, with ηt > ηtþ1 , and those based on the
period t + 1 technology are ηt and ηtþ1 , respectively, with ηt > ηtþ1 . Their average
distance ηt - ηtþ1 þ ηt - ηtþ1 =2, is thus greater than zero, which indicates
that the productivity of DMU A has increased from period t to period t + 1. Note that
ηtþ1 < 0, because At + 1 lies outside of the period t frontier.
Fig. 8.5 Measurement of Y

the LPI
ˆ t +1
Period t+1
Kˆ t +1 A
t +1
A
Period t
A t +1
At K t +1
Kt
Ât
At
Kˆ t
B C
O
X
The LPI is a difference-measure counterpart of the ratio measure of MPI. First,

the LPI is the arithmetic mean of the period t and period t + 1 LPIs, versus the
geometric mean of the two ratio-measure MPIs. Second, the LPI can be decomposed
into the sum of the efficiency change and technical change, as opposed to the product
decomposition of the MPI. Specifically, we have, from (8.18):
→t → tþ1
LPI Zt , dt , Ztþ1 , dtþ1 = D ðZt , dt Þ - D Ztþ1 , dtþ1 þ
1 → tþ1 →t → tþ1 →t
D ðZt , dt Þ - D ðZt , dt Þ þ D Ztþ1 , dtþ1 - D Ztþ1 , dtþ1
2
= ðefficiency changeÞ þ ðtechnical changeÞ
Third, similar to the MPI, the LPI may also be infeasible under the variable
returns to scale technology.
Other similar forms of LPI can also be devised from the directional distance
function. For example, Chung et al. (1997) used the ratio of two directional distance
functions, each added to one, to represent the change in productivity between two
periods. Since this index is the directional distance functions in ratio form, it is called
the Malmquist-Luenberger productivity index. This index, for a DMU between
periods t and t + 1 based on the period s technology, is:
→s
1 þ D ðZt , dt Þ
ML Z , d , Z
s t t tþ1
,d tþ1
= →s ð8:19Þ
1 þ D Ztþ1 , dtþ1
This index clearly possesses the property of circularity. If period t (or t + 1) is

used as the base period, then the index can be decomposed into the product of the
efficiency change and technical change as:
→t →t → tþ1
1 þ D ðZ t , d t Þ 1 þ D ðZt , dt Þ 1þ D Ztþ1 , dtþ1
→t
= → tþ1
× →t
1 þ D Ztþ1 , dtþ1 1þ D Ztþ1 , dtþ1 1 þ D Ztþ1 , dtþ1
Using DMU A in Fig. 8.5 to explain this, the efficiency change is the ratio of the
t tþ1
efficiency at period t + 1 to that at period t, which is A B=At B = A C=Atþ1 C ,
and the technical change is the shift of the frontier from period t to period t + 1,
tþ1 tþ1
which is A C=Atþ1 C = A C=Atþ1 C : Note again that ηtþ1 < 0: To avoid the
use of an arbitrary reference technology, the geometric mean of two indexes
measured from the two periods as the base period is used to obtain the following
result:
1=2
ML Zt , dt , Ztþ1 , dtþ1 = MLt Zt , dt , Ztþ1 , dtþ1 × MLtþ1 Zt , dt , Ztþ1 , dtþ1
→ tþ1 1=2
→t → tþ1
1 þ D ðZ t , d t Þ 1þ D Ztþ1 , dtþ1 1þ D ðZt , dt Þ
= → tþ1
× →t →t
1þ D Ztþ1 , dtþ1 1 þ D Ztþ1 , dtþ1 1 þ D ðZt , dt Þ
→t
In this setting 1 þ D ðZt , dt Þ is the inverse of the efficiency of Zt at period t, and
→ tþ1
1þ D Ztþ1 , dtþ! is that of Zt + 1 at period t + 1, such that their ratio becomes the
ratio of the period t + 1 efficiency to the period t efficiency. The technical change is
the geometric mean of the two technical changes measured from Zt and Zt + 1.
The Luenberger productivity index is generally used in cases of undesirable
outputs, when the desirable outputs are to be increased while the undesirable outputs
are to be decreased. To avoid infeasibility under variable returns to scale, a global
index that uses all DMUs at all periods can be used to construct the global frontier
(Oh 2010).
The Malmquist productivity index has been widely applied to measure productivity
changes in different areas, for example, football teams (Guzman and Morrow 2007),
forests (Kao 2010), the semiconductor industry (Lee and Johnson 2011), the services
sector (Lee 2013), railroad companies (Abate et al. 2013), power plants (Zhang and
Choi 2013), education (Brennan et al. 2014), the IT industry (Chou and Shao 2014),
banks (Kao and Liu 2014), the tourism industry (Sun et al. 2015), hospitals
(Kittelsen et al. 2015), the textile industry (Kapelko and Lansink (2015), fish farming
(Iliyasu et al. 2015), and the cement industry (Ghulam and Jaffry 2015). The global
Malmquist productivity index has been proposed, in different forms (Asmild and
Tam 2007, Johidi et al. 2012, Afsharian and Heinz 2015), in order to provide a
consistent measure of productivity change based on the technology of different
periods.
The productivity change based on the directional distance function is measured
by the Luenberger and Malmquist-Luenberger indicators. Several papers discuss the
infeasibility of these indicators (Briec and Kerstens 2009; Aparicio et al. 2013; Chen
et al. 2013). Different types of slacks-based (Arabi et al. 2015) and non-radial
Russell type (Mahlberg and Sahoo 2011, Zhang and Choi 2013) indexes have
been proposed, and there are also other studies on this type of index (Färe and
Primont 2003; Kapelko et al. 2015; Lansink et al. 2015).
References
Abate, M., Lijesen, M., Pels, E., & Roelevelt, A. (2013). The impact of reliability on the produc-
tivity of railroad companies. Transportation Research Part E- Logistics and Transportation
Review, 51, 41–49.
Afsharian, M., & Heinz, A. (2015). The overall Malmquist index: A new approach for measuring
productivity changes. Annals of Operations Research, 226, 1–27.
Aparicio, J., Pastor, J. T., & Zofio, J. L. (2013). On the inconsistency of the Malmquist-Luenberger
index. European Journal of Operational Research, 229, 738–742.
Arabi, B., Munisamy, S., & Emrouznejad, A. (2015). A new slacks-based measure of Malmquist-
Luenberger index in the presence of undesirable outputs. Omega, 51, 29–37.
Asmild, M., & Tam, F. (2007). Estimating global frontier shifts and global Malmquist indices.
Brennan, S., Haelermans, C., & Ruggiero, J. (2014). Nonparametric estimation of education
productivity incorporating nondiscretionary inputs with an application to Dutch schools.
Briec, W., & Kerstens, K. (2009). Infeasibility and directional distance functions with application to
the determinateness of the Luenberger productivity indicator. Journal of Optimization Theory
and Applications, 141, 55–73.
Caves, D. W., Christensen, L. R., & Diewert, W. E. (1982). The economic theory of index numbers
and the measurement of input, output, and productivity. Econometrica, 50, 1393–1414.
Chambers, R. G., Chung, Y., & Färe, R. (1996). Benefit and distance functions. Journal of
Chambers, R. G., & Pope, R. D. (1996). Aggregate productivity measures. American Journal of
Agricultural Economics, 78, 1360–1365.
Chang, H., Choy, H. L., Cooper, W. W., & Ruefli, T. W. (2009). Using Malmquist indexes to
measure changes in the productivity and efficiency of US accounting firms before and after the
Sarbanes-Oxley act. Omega, 37, 951–960.
Chen, Y., Du, J., & Huo, J. (2013). Super-efficiency based on a modified directional distance
function. Omega, 41, 621–625.
Chou, Y. C., & Shao, B. B. M. (2014). Total factor productivity growth in information technology
services industries: A multi-theoretical perspective. Decision Support Systems, 62, 106–118.
References 181
Chung, Y. H., Färe, R., & Grosskopf, S. (1997). Productivity and undesirable outputs: A directional
distance production approach. Journal of Environmental Management, 51, 229–240.
Färe, R., Grosskopf, S., Norris, M., & Zhang, Z. (1994). Productivity growth, technical progress,
and efficiency change in industrialized countries. American Economic Review, 84, 66–83.
Färe, R., & Primont, D. (2003). Luenberger productivity indicators: Aggregation across firms.
Ghulam, Y., & Jaffry, S. (2015). Efficiency and productivity of the cement industry: Pakistani
experience of deregulation and privatization. Omega, 54, 101–115.
Guzman, I., & Morrow, S. (2007). Measuring efficiency and productivity in professional football
teams: Evidence from the English premier league. Central European Journal of Operations
Research, 15, 309–328.
Iliyasu, A., Mohamed, Z. A., & Hashim, M. (2015). Productivity growth, technical change and
efficiency change of the Malaysian cage fish farming: An application of Malmquist productivity
index approach. Aquaculture International, 23, 1013–1024.
Johidi, G., Razavyan, S., & Tohidnia, S. (2012). A global Malmquist productivity index using data
envelopment analysis. Journal of the Operational Research Society, 63, 72–78.
Kao, C. (2000). Measuring the performance improvement of Taiwan forests after reorganization.
Forest Science, 46, 577–584.
Kao, C. (2010). Malmquist productivity index based on common-weights DEA: The case of Taiwan
forests after reorganization. Omega, 38, 484–491.
Kao, C., & Liu, S. T. (2014). Multi-period efficiency measurement in data envelopment analysis:
The case of Taiwanese commercial banks. Omega, 47, 90–98.
Kapelko, M., Horta, I. M., Camanho, A. S., & Lansink, A. O. (2015). Measurement of input-
specific productivity growth with an application to the construction industry in Spain and
Portugal. International Journal of Production Economics, 166, 64–71.
Kapelko, M., & Lansink, A. O. (2015). An international comparison of productivity change in the
textile and clothing industry: A bootstrapped Malmquist index approach. Empirical Economics,
48, 1499–1523.
Kittelsen, S. A. C., Winsnes, B. A., Anthun, K. S., Goude, F., Hope, O., Hakkinen, U., Kalseth, B.,
Kilsmark, J., Medin, E., Rehnberg, C., & Ratto, H. (2015). Decomposing the productivity
differences between hospitals in the Nordic countries. Journal of Productivity Analysis, 43,
281–293.
Lansink, A. O., Stefanou, S., & Serra, T. (2015). Primal and dual dynamic Luenberger productivity
indicators. European Journal of Operational Research, 241, 555–563.
Lee, B. L. (2013). Productivity, technical and efficiency change in Singapore’s services sector, 2005
to 2008. Applied Economics, 45, 2023–2029.
Lee, C. Y., & Johnson, A. L. (2011). A decomposition of productivity change in the semiconductor
manufacturing industry. International Journal of Production Research, 49, 4761–4785.
Luenberger, D. G. (1992). Benefit functions and duality. Journal of Mathematical Economics, 21,
461–481.
Mahlberg, B., & Sahoo, B. K. (2011). Radial and non-radial decompositions of Luenberger
productivity indicator with an illustrative application. International Journal of Production
Economics, 131, 721–726.
Malmquist, S. (1953). Index numbers and indifference curves. Trabajos de Estatistica, 4, 209–242.
Oh, D. H. (2010). A global Malmquist-Luenberger productivity index. Journal of Productivity
Analysis, 34, 183–197.
Pastor, J. T., & Lovell, C. A. K. (2005). A global Malmquist productivity index. Economics Letters,
88, 266–271.
Pastor, J. T., & Lovell, C. A. K. (2007). Circularity of the Malmquist productivity index. Economic
Theory, 33, 591–599.
Sun, J. R., Zhang, J., Zhang, J. H., Ma, J. H., & Zhang, Y. L. (2015). Total factor productivity
assessment of tourism industry: Evidence from China. Asia Pacific Journal of Tourism
Research, 20, 280–294.
Zhang, N., & Choi, Y. (2013). Total factor carbon emission performance of fossil fuel power plants
in China: A metafrontier non-radial Malmquist index analysis. Energy Economics, 40, 549–559.
Chapter 9
Basic Ideas in Efficiency Measurement
for Network Systems
In measuring the efficiency of a DMU one can consider it as a closed plant, and the
evaluator stands outside of the plant counting the quantity of materials and number
of workers entering to manufacture products. After a period of time, the evaluator
counts the quantity of products sent out of the plant. From the inputs consumed and
the outputs produced, the evaluator is able to measure the relative efficiency of this
plant compared to other similar plants via the conventional DEA technique. The
plant in this case is treated as a black box, in that how the materials are converted into
outputs inside the plant is not known.
Consider a case of six plants, each uses inputs X1 and X2 to produce outputs Y1
and Y2, with the data shown in columns two to five of Table 9.1. Based on the CCR
model, the efficiencies of the six plants are calculated, as shown in column six, where
only plant F is inefficient. Being aware that it is the only inefficient plant, the
manager of plant F decides to make improvements and asks the evaluator for help.
To examine how much of the inputs can be reduced and how much of the outputs can
be increased under the current technology, the evaluator enters the plant and finds
that the whole production process includes two major operations performed at two
workstations. Specifically, station 1 applies inputs X1 and X2 to produce an interme-
diate product Z, and station 2 uses the intermediate product to produce the final
outputs Y1 and Y2, as depicted in Fig. 9.1.
The evaluator thus collects the data for the intermediate product of the six plants,
as shown in column seven of Table 9.1. By treating station 1 as an independent
DMU, with X1 and X2 as the inputs and the intermediate product Z as the output, the
efficiency of station 1 is measured. Similarly, by treating station 2 as an independent
DMU, with the intermediate product Z as the input and Y1 and Y2 as the outputs, the
efficiency of station 2 is also measured. The last two columns of Table 9.1 show the
efficiencies of the two stations measured from the CCR model. The evaluator then
notices that when each station is treated as an independent DMU, none of the five
efficient plants, A to E, is efficient in both stations. More seriously, plants A and
https://doi.org/10.1007/978-3-031-27593-7_9
184 9 Basic Ideas in Efficiency Measurement for Network Systems
Table 9.1 Data and efficiency measures for an example

Inputs Outputs Station efficiency
Plant X1 X2 Y1 Y2 Plant efficiency Intermediate product Z Station 1 Station 2
A 1 2 2 3 1 3 0.75 0.72
B 1 2 3 2 1 4 1 0.54
C 2 1 2 3 1 4 1 0.54
D 2 1 3 2 1 3 0.75 0.72
E 2 4 5 5 1 3.6 0.45 1
F 2 3 3 2 0.6 3 0.45 0.72
Fig. 9.1 A plant with two Plant

workstations
Z
X1, X2 Station 1 Station 2 Y1, Y2
D are inefficient in both stations, yet are evaluated as efficient ones when the plant is
treated as a whole unit.
This example shows that the efficiency of the plant, as a whole unit, will be
overstated if the operations of the stations, as the components of the whole unit, are
not considered in efficiency measurement. The overstatement is expected to be more
serious when more stations are involved. Thus, whenever the data is available, the
operations of the stations should be considered to obtain meaningful results.
A plant is usually composed of a number of stations working interdependently, in
that a station produces some intermediate products for other stations to use to form a
network production system. There are different ways of accounting for the opera-
tions of the stations in measuring the efficiency of the plant. Kao (2009) classified the
associated models into three types: independent, connected, and relational. Kao
(2018) separated another type of models, called cooperative type, from the relational
type. They will be introduced in Sects. 9.2, 9.3, 9.4, and 9.5, respectively. Different
terms have been used in the literature when measuring the efficiency of network
systems, such as DMU and system for the whole unit of concern, and sub-DMU,
sub-unit, sub-process, and division for the components of the whole unit of concern.
In this book we use DMU and system interchangeably to refer to the whole unit of
concern and division to refer to a component of the DMU when there is no
ambiguity. These four types of models are based on the ratio measure, distance
function measure, and slacks-based measures of efficiency discussed in Chaps. 2, 3,
and 4, respectively. Before getting into the detail of how these models are applied to
measure the efficiency of a network system when the operations of the divisions are
considered, we explain in the next section how the four efficiency measures are
9.1 The Black-Box Model 185
applied to the network system when the operations of the divisions are not consid-
ered as a basis for comparison.
9.1 The Black-Box Model
ðk Þ ðk Þ
Consider a system composed of p divisions. Denote X ij and Y rj as the ith input,
i = 1, . . ., m, supplied from outside to and the rth final output, r = 1, . . ., s, produced
from the kth division, k = 1, . . ., p, of the jth DMU, j = 1, . . ., n, respectively.
Following the conventional notation of Xij for the ith input used and Yrj for the rth
ðk Þ ðk Þ
output produced by the jth DMU, we have X ij = pk = 1 X ij and Y rj = pk = 1 Y rj .
One feature that characterizes the network system is the intermediate product.
Different from the exogenous inputs that are supplied from outside and the final
outputs that are produced for outside, the intermediate products are produced and
ða,bÞ
consumed within the system, and thus are not visible from outside. Let Z gj denote
the gth intermediate product, g = 1, . . ., h, produced by division a for division b to
ðk,bÞ
use. pb = 1 Z gj is then the total amount of the gth intermediate product produced by
ða,k Þ
division k for other divisions to use, and pa = 1 Z fj is the total amount of the fth
intermediate product produced by other divisions for division k to use. Theoretically,
every division can consume all inputs and intermediate products and produce all
outputs and intermediate products. In reality, however, a division will consume only
certain inputs and intermediate products and produce certain outputs and intermedi-
ðk Þ ðk Þ ða,bÞ
ate products. In other words, many X ij , Y rj , and Z gj are zero. Figure 9.2 shows a
general structure for network systems.
Suppose the system is treated as a black box, in that only the exogenous inputs
supplied from outside and the final outputs produced for outside are considered. The
CCR input model in multiplier form is:
Fig. 9.2 General network

systems
s p
ðkÞ
max: ur Y r0
r=1 k=1
m p
ðkÞ
s:t: vi X i0 =1
i=1 k=1
ð9:1Þ
s p m p
ðkÞ ðkÞ
ur Yr j - vi Xi j ≤ 0, j = 1, . . . , n
r=1 k=1 i=1 k=1
ur , vi ≥ 0, r = 1, . . . , s, i = 1, . . . , m
Here, the non-Archimedean number ε as a lower bound for the multipliers is

ignored to make the expression simpler.
The CCR input model in envelopment form is:
min:θ
n p p
ðkÞ ðkÞ
s:t: λj Xi j ≤θ X i0 , i = 1, . . . , m
j=1 k=1 k=1
n p p
ð9:2Þ
ðkÞ ðkÞ
λj Yr j ≥ Y r0 , r = 1, . . . , s
j=1 k=1 k=1
λj ≥ 0, j = 1, . . . , n
Since this model is the dual of Model (9.1), it has the same objective value as that
of Model (9.1).
The slacks-based model is:
1 m p ðkÞ
1- i=1
si- = k=1
X i0
min: m
1 s p ðkÞ
1þ r=1
sþ
r = k=1
Y r0
s
n p p
ðkÞ ðkÞ
s:t: λj Xi j þ si- = X i0 , i = 1, . . . , m ð9:3Þ
j=1 k=1 k=1
n p p
ðkÞ ðkÞ
λj Yr j - sþ
r = Y r0 , r = 1, . . . , s
j=1 k=1 k=1
λj , si , sþ
-
r ≥ 0, 8j, i, r
The constraints of this model are essentially the same as those of Model (9.2).
These two models differ only in how they define efficiency.
These three DMU-based models have not considered the operations of the
divisions, so their efficiency measures will be overstated. They serve as a basis for
the three types of models to be discussed in the succeeding sections for comparison
purposes.
9.2 Independent Model 187
9.2 Independent Model
The most straightforward way to investigate the performance of a network system is

to measure the efficiency of each division by treating them as independent DMUs.
That is, every division of a DMU is compared with the corresponding division of
other DMUs by applying the conventional DEA models. There are three forms of the
independent model: multiplier, envelopment, and slacks-based, as detailed below.
9.2.1 Multiplier Form
Suppose the CCR input model in multiplier form is applied. The efficiency, for
example, of the kth division of a DMU is measured as:
s h p
ðkÞ ðkÞ ðkÞ ðk,bÞ
max: ur Y r0 þ wg Z g0
r=1 g=1 b=1
m h p
ðkÞ ðkÞ ðkÞ ða,kÞ
s:t: vi X i0 þ ^f
w Zf 0 =1
i=1 f =1 a=1
s h p
ðkÞ ðkÞ
ur Y rj þ wg
ðkÞ
Z gj
ðk,bÞ
- ð9:4Þ
r=1 g=1 b=1
m h p
ðkÞ ðkÞ ðkÞ ða,kÞ
vi X ij þ ^f
w Zf j ≤ 0, j = 1, . . . , n
i=1 f =1 a=1
ðkÞ ðkÞ ðkÞ ðkÞ

^ f ≥ 0, 8r, i, g, f
ur , vi , wg , w
In this model the intermediate inputs consumed by and the intermediate outputs
produced from the division are also considered in measuring its efficiency. The
results show how the kth division has performed as compared to those of other
DMUs. The independent model inherits the property of the CCR model that the input
and output models yield the same efficiency measure.
Model (9.4) corresponds to one division of each DMU. By solving Model (9.4)
for all p divisions, one at a time, the efficiencies of all p divisions are obtained. All
these p linear programs can be aggregated to form a separable program of the
following form, such that the efficiencies of all p divisions can be obtained at
one time:
p s h p
ðk Þ ðk,bÞ
max: uðrkÞ Y r0 þ wðgkÞ Z g0
k=1 r=1 g=1 b=1
m h p
ðk Þ ðk Þ ðk Þ ða,k Þ
s:t: vi X i0 þ wf Zf 0 = 1, k = 1, . . . , p
i=1 f =1 a=1
s h p
ðk Þ ðk,bÞ
uðrkÞ Y rj þ wðgkÞ Z gj -
r=1 g=1 b=1
m h p
vi X ij þ wf Z fj ≤ 0, j = 1, . . . , n, k = 1, . . . , p
i=1 f =1 a=1
ðk Þ ðk Þ
uðrkÞ , vi , wðgkÞ , wf ≥ 0, 8r, i, g, f , k
ð9:5Þ
Each term in square brackets in the objective function corresponds to the effi-
ciency of a division, and the constraints are those of Model (9.4) enumerated for
p divisions.
9.2.2 Envelopment Form
Suppose the CCR input model in envelopment form is used to measure the division
efficiency. The model will be the dual of Model (9.4), which is:
min:θðkÞ
n
ðkÞ ðkÞ ðkÞ
s:t: λj X i j ≤ θðkÞ X i0 , i = 1, . . . , m
j=1
n p p
ðkÞ ða, kÞ ða, kÞ
λj Zf j ≤ θðkÞ Zf 0 , f = 1, . . . , h
j=1 a=1 a=1
n p p ð9:6Þ
ðkÞ ðk, bÞ ðk, bÞ
λj Zg j ≥ Z g0 , g = 1, . . . , h
j=1 b=1 b=1
n
ðkÞ ðkÞ ðkÞ
λj Y r j ≥ Y r0 , r = 1, . . . , s
j=1
ðkÞ
λj ≥ 0, 8j
The CCR model has two forms: input- and output-oriented. Since they yield the
same efficiency measure, the output-oriented form is not formulated here.
Similar to the case of the multiplier form, the models for all p divisions can be
aggregated to form the following separable program, where the efficiencies of all
p divisions are measured in one model:
p
min: θðkÞ
k=1
n
ðkÞ ðkÞ ðkÞ
s:t: λj X i j ≤ θðkÞ X i0 , i = 1, . . . , m, k = 1, . . . , p
j=1
n p p
λj Zf j ≤ θðkÞ Zf 0 , f = 1, . . . , h, k = 1, . . . , p
j=1 a=1 a=1 ð9:7Þ
n p p
ðkÞ ðk , bÞ ðk , bÞ
λj Zg j ≥ Z g0 , g = 1, . . . , h, k = 1, . . . , p
j=1 b=1 b=1
n
ðkÞ ðkÞ ðkÞ
λj Y r j ≥ Y r0 , r = 1, . . . , s, k = 1, . . . , p
j=1
ðkÞ
λj ≥ 0, 8j, k
The objective function is the sum of the distance parameters corresponding to

p divisions, and the constraints are those of Model (9.6) enumerated for p divisions.
At optimality, θ(k) is the efficiency of the kth division.
9.2.3 Slacks-Based Form
The efficiency of the division can also be measured using a slacks-based measure
(SBM) model of the following form:
1 m ðk Þ - ðk Þ h ðk Þ - p ða,k Þ
1- si =X i0 þ sf = Z
min: mþh i=1 f =1 a=1 f0
1 s ðk Þ h p ðk,bÞ
1þ sðrkÞþ =Y r0 þ sðgkÞþ = Z
sþh r=1 g=1 b = 1 g0
n
ðk Þ ðk Þ ðk Þ - ðk Þ
s:t: λj X ij þ si = X i0 , i = 1, . . . , m
j=1
n p p
ðk Þ ða,k Þ ðk Þ - ða,k Þ
λj Z fj þ sf = Zf 0 , f = 1, . . . , h, ð9:8Þ
j=1 a=1 a=1
n p p
ðk Þ ðk,bÞ ðk,bÞ
λj Z gj - sðgkÞþ = Z g0 , g = 1, . . . , h,
j=1 b=1 b=1
n
ðk Þ ðk Þ ðk Þ
λj Y rj - sðrkÞþ = Y r0 , r = 1, . . . , s,
j=1
ðk Þ ðk Þ - ðk Þ -
λj , s i , sf , sðgkÞþ , sðrkÞþ ≥ 0, 8j, i, f , g, r
This model is nonlinear, which can be linearized via a variable substitution

technique (Charnes and Cooper (1962)) as discussed in Chap. 4. The model for
each division can be aggregated to form the following separable program:
1 m ðkÞ- ðk Þ h ðkÞ- p ða,k Þ

p 1- si =X i0 þ sf = Z
min: mþh i=1 f =1 a=1 f0
1 s ðk Þ h p ðk,bÞ
k=1 1þ sðrkÞþ =Y r0 þ sðgkÞþ = Z
sþh r=1 g=1 b = 1 g0
n
ðk Þ ðk Þ ðkÞ- ðk Þ
s:t: λj X ij þ si = X i0 , i = 1, . . . , m, k = 1, . . . , p
j=1
n p p
ðk Þ ða,k Þ ðk Þ- ða,k Þ
λj Z fj þ sf = Zf 0 , f = 1, . . . , h, k = 1, . . . , p
j=1 a=1 a=1
n p p
λj Z gj - sðgkÞþ = Z g0 , g = 1, . . . , h, k = 1, . . . , p
j=1 b=1 b=1
n
λj Y rj - sðrkÞþ = Y r0 , r = 1, . . . , s, k = 1, . . . , p
j=1
ðk Þ ðk Þ- ðk Þ-
λj , si , sf , sðgkÞþ , sðrkÞþ ≥ 0, 8j, i, f , g, r, k
ð9:9Þ
Each ratio in parentheses in the objective function is the efficiency of a division,

and the constraints are those of Model (9.8) enumerated for p divisions.
Other ways for measuring the efficiency, e.g. the use of directional distance
measures, and other models, such as output-oriented and variable returns to scale,
can also be applied to measure the efficiency of each division. These approaches
should be straightforward if readers are familiar with the models introduced in
Chaps. 2, 3, and 4.
The independent model emphasizes the efficiency of the divisions. If the effi-
ciency of a DMU is desired, then it must be defined as an aggregation of the division
efficiencies and expressed as the objective function. Different definitions of system
efficiency have appeared in the literature, and these will be discussed in the related
sections in the succeeding chapters.
9.3 Connected Model
The previous example illustrates that the system efficiency will be overstated if the
operations of the divisions are ignored. Färe and Grosskopf (1996a, 2000) formu-
lated the production possibility set for a network system and used the distance
functions defined on the production possibility set as the efficiency of the system.
A large number of network DEA studies use this idea to develop models to measure
9.3 Connected Model 191
the efficiency of the system. In the modeling divisions are connected via the
intermediate products, and the associated models are thus named the connected
model by Kao (2009). The connected model also has three forms: envelopment,
multiplier, and slacks-based.
For a general network system with the structure shown in Fig. 9.2, the corresponding
production possibility set under constant returns to scale is:
p
^ j
n n
T = ðX, Y, Z, ZÞ ðkÞ ðkÞ
λj X ij ≤ X i
ðkÞ
, λj
ðkÞ ða,kÞ
Zf j
ðkÞ
≤ Z^ f ,
j=1 j=1 a=1
n p n
ðkÞ ðk,bÞ ðkÞ ðkÞ ðkÞ
λj Z gj ≥ Z gðkÞ , λj Y rj ≥ Y rðkÞ , λj ≥ 0, 8i, f , g, r, j, k
j=1 b=1 j=1
ð9:10Þ
If division k has a variable returns to scale technology, then the constraint

n ðk Þ
j = 1 λj= 1 is added. The connected model is usually formulated in envelopment
form. From the input side, the model for measuring the system efficiency is:
min: θ
n
ðkÞ ðkÞ ðkÞ
s:t: λj X i j ≤ θX i0 , i = 1, . . . , m, k = 1, . . . , p
j=1
n p p
λj Zf j ≤ Zf 0 , f = 1, . . . , h, k = 1, . . . , p
j=1 a=1 a=1
n p p ð9:11Þ
λj Zg j ≥ Z g0 , g = 1, . . . , h, k = 1, . . . , p
j=1 b=1 b=1
n
ðkÞ ðkÞ ðkÞ
λj Y r j ≥ Y r0 , r = 1, . . . , s, k = 1, . . . , p
j=1
ðkÞ
λj ≥ 0, 8j, k
Compared with the corresponding black-box model (9.2), it is noted that this
model also has constraints corresponding to the intermediate products. Moreover,
the first and fourth constraint sets are the division separations of the two sets of
ðk Þ
constraints of Model (9.2). In the separation, the linear combination variables λj are
allowed to be different for different divisions. On the one hand, Model (9.11) has
more constraints, which makes it more stringent than Model (9.2), and on the other
hand it has more variables, which makes it less stringent. Whether the aggregate
effect is more or less stringent with regard to having a higher or lower efficiency
depends on the structure of the problem.
Compared to the corresponding independent model (9.6), there are two differ-
ences to be noted. One is that the distance parameter θ is the same for all p divisions,
and the other is that θ is only attached to the exogenous inputs. Model (9.11) is thus
more stringent, although their objective functions are different.
A merit of the envelopment model is it shows the projection points of the factors
for inefficient DMUs to make improvements. The factor with a caret “^” is the
corresponding projection point. The constraints of Model (9.11) show the projection
ðk Þ p ða,k Þ p ðk,bÞ ðk Þ
points for X i0 , a = 1Zf 0 , b = 1 Z g0 , and Y r0 as:
n
^ ðk Þ =
X
ðk Þ
λ j X ij ,
ðk Þ
i = 1, . . . , m, k = 1, . . . , p
i0
j=1
n p
ðk ÞðinÞ ðk Þ ða;k Þ
Z^ f 0 = λj Z fj , f = 1, . . . , h, k = 1, . . . , p
j=1 a=1
n p
ðk ÞðoutÞ ðk Þ ðk;bÞ
Z^ g0 = λj Z gj g = 1, . . . , h, k = 1, . . . , p
j=1 b=1
n
Y^ r0 = λ j Y rj , r = 1, . . . , s, k = 1, . . . , p
j=1
ðk ÞðinÞ
where Z f 0 is the total amount of intermediate product f flowing into division k,
ðk ÞðoutÞ
and Z g0 is the total amount of intermediate product g flowing out of division k.
The projection points serve as targets for the corresponding factors to achieve, so that
the inefficient DMU will become efficient.
The connected model in envelopment form can also be formulated from the
output side as:
max: φ
n
ðkÞ ðkÞ ðkÞ
s:t: λj X i j ≤ X i0 , i = 1, . . . , m, k = 1, . . . , p
j=1
n p p
λj Zf j ≤ Zf 0 f = 1, . . . , h, k = 1, . . . , p
j=1 a=1 a=1
n p p ð9:12Þ
λj Zg j ≥ Z g0 , g = 1, . . . , h, k = 1, . . . , p
j=1 b=1 b=1
n
ðkÞ ðkÞ ðkÞ
λj Y r j ≥ φY r0 , r = 1, . . . , s, k = 1, . . . , p
j=1
ðkÞ
λj ≥ 0, 8j, k
9.3 Connected Model 193
The system efficiency is the inverse of φ. In the black-box system the input
efficiency is equal to the output efficiency under constant returns to scale. In the
network system whether the input efficiency θ measured from Model (9.11) is the
same as the output efficiency 1/φ measured from this model is not known. However,
this can be judged from the multiplier form of the model discussed in the following
subsection.
The multiplier form of a model is the dual of the envelopment form. Based on Model
(9.11), the input-oriented model in multiplier form is:
p s p h p p h p
ðk Þ ðk,bÞ ðk Þ ða,k Þ
max: uðrkÞ Y r0 þ wðgkÞ Z g0 - wf Zf 0
k=1 r=1 k=1 g=1 b=1 k=1 f =1 a=1
p m
ðk Þ ðk Þ
s:t: vi X i0 = 1
k=1 i=1
s h p
ðk Þ ðk,bÞ
uðrkÞ Y rj þ wðgkÞ Z gj -
r=1 g=1 b=1
m h p
i=1 f =1 a=1
ðk Þ ðk Þ
ð9:13Þ
This model is very similar to its counterpart of independent model (9.5), only the
normalization constraint is the sum of the p normalization constraints of Model (9.5),
with the aggregate intermediate input moved to the objective function. The input
efficiency is
p s ðk Þ ðk Þ h ðk Þ p ðk,bÞ h ðk Þ p ða,kÞ
k=1 u
r=1 r Y r0 þ g=1 gw Z
b = 1 g0 - w
f =1 f Z
a=1 f0 /
p m ðk Þ ðkÞ
k=1 i = 1 vi X i0 .
The multiplier form of the output-oriented model is the dual of Model (9.12),
which is:
p m p h p p h p
ðk Þ ðk Þ ðk Þ ða,k Þ ðk,bÞ
min: vi X i0 þ wf Zf 0 - wðgkÞ Z g0
k=1 i=1 k=1 f =1 a=1 k=1 g=1 b=1
p s
ðk Þ
s:t: uðrkÞ Y r0 = 1
k=1 r=1
m h p
vi X ij þ wf Z fj -
i=1 f =1 a=1
s h p
ðk Þ ðk,bÞ
uðrkÞ Y rj þ wðgkÞ Z gj ≥0 j = 1, . . . , n, k = 1, . . . , p
r=1 g=1 b=1
ðk Þ ðk Þ
ð9:14Þ
p s ðk Þ ðk Þ p m ðk Þ ðk Þ
The output efficiency is k=1 r = 1 ur Y r0 = k=1 i = 1 vi X i0
p h ðk Þ p ða,k Þ p h ðk Þ p ðk,bÞ
+ k=1 f = 1 wf ð a = 1Z f 0 Þ - k=1 g = 1 wg b = 1 Z g0 : This will be
the same as the input efficiency only if the aggregate intermediate input is equal to
p h ðk Þ p ða,kÞ
the aggregate intermediate output, i.e., k=1 f = 1 wf a = 1Zf 0 =
p h ðk Þ p ðk,bÞ
k=1 g = 1 wg b = 1 Z g0 : Since this is seldom the case, the input efficiency
under constant returns to scale in general differs from the output efficiency for
network systems.
In Fig. 9.2 it is assumed that all outputs Y ðrkÞ of a division are different from all of
its intermediate outputs for simplicity of notation. In the real world, intermediate
products may also be final outputs, as in the case of spare parts. Fig. 9.2 also assumes
the exogenous inputs supplied from outside to a division to be different from the
intermediate inputs supplied by other divisions. In practice it is also possible that an
input needed by a division be either supplied from outside or another division in the
system. In this case an exogenous input of a division is the same as one of its
intermediate inputs. We will use an example in Sect. 9.6 to explain how to construct
the model in these cases.
The SBM efficiency of the connected model, based on the idea of Tone and Tsutsui
(2009), can be formulated as:
9.4 Relational Model 195
p 1 m ðk Þ - ðk Þ
k=1
1- i=1
si =X i0
min: m
p 1 s ðk Þ
k=1
1þ r=1
sðrkÞþ =Y r0
s
n
s:t: λj X ij þ si = X i0 , i = 1, . . . , m, k = 1, . . . , p
j=1
n p p
λj Z fj þ sf = Zf 0 , f = 1, . . . , h, k = 1, . . . , p
j=1 a=1 a=1
n p p
λj Z gj - sðgkÞþ = Z g0 , g = 1, . . . , h, k = 1, . . . , p
j=1 b=1 b=1
n
λj Y rj - sðrkÞþ = Y r0 , r = 1, . . . , s, k = 1, . . . , p
j=1
λj , s i , sf , sðgkÞþ , sðrkÞþ ≥ 0, 8j, i, f , g, r, k
ð9:15Þ
Since the connected model is used to measure the system efficiency while taking
the operations of the divisions into account, the constraints are the same as its
independent model counterpart of Model (9.9). The objective function differs from
that of Model (9.9) in that only the slacks related to the exogenous inputs and final
outputs are considered. Compared to the envelopment form of either the input model
(9.11) or output model (9.12), it is noted that their constraints are essentially the
same, and they differ only in the way that the efficiency is defined.
Similar to the envelopment form of the connected model, the slacks-based
approach is also able to show the projection points of the factors for inefficient
DMUs to become efficient.
9.4 Relational Model
The independent model measures the division efficiencies independently. If the

system efficiency is desired, then it must be defined externally, usually as an
aggregation of the division efficiencies. The connected model measures the system
efficiency, taking the operations of the divisions into consideration. Usually the
efficiencies of the divisions are not obtainable, and they must be measured via an
independent model by treating the divisions as DMUs. Combining the concepts of
these two models, Kao (2009) proposed a relational model, which is able to measure
the efficiency of both the system and the divisions. Moreover, a relationship between
the system efficiency and the division efficiencies can be derived for different
network structures.
Similar to the connected model, the major objective of the relational model is to
measure the system efficiency. However, the relational model is able to measure the
efficiencies of the divisions at the same time without extra effort, when the opera-
tions of the divisions are taken into account. The relational model is formulated in
multiplier form, with the characteristic that the same factor has the same multiplier,
no matter which division it corresponds to, or what roles, e.g. input, output, or
intermediate product, it plays. Conceptually, it is the black-box model, with the
operations of the divisions additionally considered, which is expressed as the
aggregate output of the division being no greater than its aggregate input. Based
on the network shown in Fig. 9.2, the input model under constant returns to scale is:
p s
ðkÞ
max: ur Y r0
k = 1r = 1
p m
ðkÞ
s:t: vi X i0 = 1
k = 1i = 1
p s p m
ðkÞ ðkÞ
ur Y r j - vi X i j ≤ 0, j = 1, . . . , n
k = 1r = 1 k = 1i = 1
s h p
ðkÞ ðk, bÞ
ur Y r j þ wg Zg j -
r=1 g=1 b=1
m h p
ðkÞ ða, kÞ
vi X i j þ wf Zf j ≤ 0, j = 1, . . . , n, k = 1, . . . , p
i=1 f =1 a=1
ur , vi , wg ≥ 0, 8r, i, g
ð9:16Þ
The objective function is the system efficiency. Each constraint in the second and
third constraint sets corresponds to one DMU and one division, respectively. Since
any intermediate product Z ðga,bÞ produced by a division a will be used by another
division b, the sum of the intermediate inputs supplied to all p divisions is equal to
the sum of the intermediate outputs produced by all p divisions, that is,
p h p ða,kÞ ðk,bÞ
k=1 f = 1 wf a = 1Zf j = pk = 1 h
w
g=1 g
p
Z
b = 1 gj
: The sum of
the p constraints corresponding to the p divisions of a DMU in the third constraint
set is thus equal to the constraint corresponding to this DMU in the second constraint
set. The second constraint set becomes redundant and can be deleted. Model (9.16) is
input-oriented, but it is clear that the output-oriented model will yield the same
efficiency measure.
Comparing this model with its black-box counterpart of Model (9.1), it is found
that these two models are the same, except that Model (9.16) has a set of constraints
9.4 Relational Model 197
corresponding to each division of all DMUs. The system efficiency calculated from
Model (9.16) is thus less than or equal to that calculated from Model (9.1). Com-
paring this model with its counterpart connected model (9.13), it is noted that Model
(9.13) boils down to Model (9.16) when the multipliers corresponding to the same
factor of different divisions are required to be the same. This implies that Model
(9.16) is more stringent, and it will thus have a smaller, or at most the same, system
efficiency.
The most important feature of Model (9.16) is that at optimality the system
efficiency, E0, is obtained from the objective value, and the division efficiencies,
ðk Þ
E 0 , are obtainable from the constraints, as follows:
p s p m p s
E0 ¼ ur Y r0 = vi X i0 ¼ ur Y r0
k¼1 r¼1 k¼1 i¼1 k¼1 r¼1
s h p m h p
ðk Þ ðk Þ ðk,bÞ ðk Þ ða,kÞ
E0 ¼ ur Y r0 þ wg Z g0 = vi X i0 þ wf Zf 0 ,
r¼1 g¼1 b¼1 i¼1 f ¼1 a¼1
k ¼ 1, . . . , p
ð9:17Þ
Under most network structures, the system efficiency can be decomposed as a

function of the division efficiencies. Kao (2009) demonstrated that there is always a
relationship between the system efficiency and the division efficiencies, after a
transformation of the network structure. This is why this type of model is called
the relational model.
If the relational model is formulated from the output side, then the objective
p m ðk Þ
function becomes min. k=1 i = 1 vi X i0 , with the normalization constraint of
p s ðk Þ
k=1 r = 1 ur Y r0 = 1 and the constraints corresponding to the divisions in Model
(9.16). In this case it is clear that the output efficiency calculated from this model is
the same as the input efficiency calculated from Model (9.16), a property possessed
by the black-box model.
The relational model can also be formulated in envelopment form. However, it is not
so straightforward to formulate it directly based on the structure shown in Fig. 9.2. It
is easier to formulate it as a dual of the multiplier form, with the redundant DMU
constraints deleted. For example, the dual of the input model (9.16) is:
min:θ
p n p
s:t: λj X ij ≤ θ X i0 , i = 1, . . . , m
k=1 j=1 k=1
p n p p
ðk Þ ðk,bÞ ða,k Þ
λj Z gj - Z gj ≥ 0, g = 1, . . . , h ð9:18Þ
k=1 j=1 b=1 a=1
p n p
λj Y rj ≥ Y r0 , r = 1, . . . , s
k=1 j=1 k=1
ðk Þ
λj ≥ 0, 8j, k
This model is different from the black-box model (9.2), in that it allows the linear
ðk Þ
combination variable λj to be different for each division and additionally considers
the constraints corresponding to the intermediate products. Compared to its coun-
terpart connected model (9.11), there are two differences to be noted. The first is that
each constraint of this model is the summation of the constraints over p divisions of
the corresponding factor. The second is that the second and third constraint sets of
Model (9.11) have been combined based on the transitivity relationship via
p p ða,kÞ ðk,bÞ
k=1 a = 1 Z fj = pk = 1 pb = 1 Z gj for f = g. Model (9.18) is thus less stringent
than Model (9.11), and consequently it has a smaller or the same system efficiency.
Since the same factors of different divisions have been combined, this model is
only able to show the projection point of each factor in the aggregate form. That is,
p p n
X i0 = X i0 = λj X ij , i = 1, . . . , m
k=1 k=1 j=1
p p n p
ðinÞ ðk ÞðinÞ ðk Þ ða,k Þ
Zf 0 = Zf 0 = λj Z fj , f = 1, . . . , h
k=1 k=1 j=1 a=1
p p n p
ðoutÞ ðkÞðoutÞ ðk Þ ðk,bÞ
Z g0 = Z g0 = λj Z gj , g = 1, . . . , h
k=1 k=1 j=1 b=1
p p n
Y r0 = Y r0 = λj Y rj , r = 1, . . . , s
k=1 k=1 j=1
There are various ways to allocate the aggregation of each factor to their com-
ponent divisions; however, they are not specified in this model.
9.5 Cooperative Model 199
The SBM efficiency of the relational model is also difficult to directly formulate.
Conceptually, its constraints are the same as those of the envelopment form of the
relational model, Model (9.18), and the objective function is related to the slacks
corresponding to the inputs and outputs of the DMU. Thus, we have
1 m ðk Þp
1- i=1
si- =X
k = 1 i0
min: m
1 s p ðk Þ
1þ r=1 r
sþ = k = 1 Y r0
s
p n p
s:t: λj X ij þ si- = X i0 , i = 1, . . . , m
k=1 j=1 k=1
p n p p n p
λj Z gj ≥ λj Z gj , g = 1, . . . , h
k=1 j=1 b=1 k=1 j=1 a=1
p n p
λj Y rj - sþ
r = Y r0 , r = 1, . . . , s
k=1 j=1 k=1
ðk Þ
λj , si- , sþ
r ≥ 0, 8j, i, g, r, k
ð9:19Þ
Similar to the case of the envelopment model (9.18), the constraints of Model
(9.19) are less stringent than those of Model (9.15), its counterpart connected model.
ðk Þ
When every division of the DMU being evaluated has the same inputs X i0 and
ðk Þ
outputs Y r0 , the objective function of Model (9.15) is the same as that of Model
(9.19). From this perspective, these two objective functions are quite similar. The
system efficiency measured from Model (9.19) is thus less than or equal to that
measured from Model (9.15).
9.5 Cooperative Model
The total amount of each intermediate product produced by all p divisions in the
system in the relational model is required to be greater than or equal to that
consumed by all p divisions in the system, as expressed by the constraint
p ðk Þ p ðk,bÞ ðk Þ ða,k Þ
k=1
n
j = 1 λj b = 1 Z gj ≥ pk = 1 nj = 1 λj p
a = 1 Z gj in Model (9.19).
This is a reasonable and necessary requirement; however, it may not be an optimal
one because when the former is strictly greater than the latter, then some amount of
this intermediate product will be wasted in the system, without being used. A
reasonable condition is to require all divisions to cooperate with each other such
that all of the intermediate products produced in the system are used up, without any
amount left in the system. This model was named cooperative model in Kao (2018).
Since it is a special case of the relational model, it shares the properties of the
relational model.
The idea of the cooperative model is stemmed from the relational model in envel-
opment form which requires the total amount of every intermediate product pro-
duced in the system to be used up in the system. Following the relational model
(9.18), only the constraints corresponding to the intermediate products are changed
from inequality to equality:
min:θ
p n p
s:t: λj X ij ≤ θ X i0 , i = 1, . . . , m
k=1 j=1 k=1
p n p p
ðk Þ ðk,bÞ ða,k Þ
λj Z gj - Z gj = 0, g = 1, . . . , h ð9:20Þ
k=1 j=1 b=1 a=1
p n p
λj Y rj ≥ Y r0 , r = 1, . . . , s
k=1 j=1 k=1
ðk Þ
λj ≥ 0, 8j, k
Since the equality constraint is stronger than the greater-than-or-equal-to con-

straint, the system efficiency calculated from the cooperative model is greater than or
equal to that calculated from the relational model.
ðinÞ ðoutÞ
The projection point (X i0 , Z g0 , Z g0 , Y r0 ) is calculated the same as that in the
ðinÞ ðoutÞ
relational model. The only difference is that Z g0 and Z g0 are the same in the
cooperative model.
The multiplier form model is the dual of the envelopment form model (9.20) which
can be expressed as:
9.5 Cooperative Model 201
p s
ðkÞ
max: ur Y r0
k = 1r = 1
p m
ðkÞ
s:t: vi X i0 = 1
k = 1i = 1
p s p m
ðkÞ ðkÞ
ur Y r j - vi X i j ≤ 0, j = 1, . . . , n
k = 1r = 1 k = 1i = 1
s h p
ðkÞ ðk, bÞ
ur Y r j þ wg Zg j -
r=1 g=1 b=1
m h p
ðkÞ ða, kÞ
vi X i j þ wf Zf j ≤ 0, j = 1, . . . , n, k = 1, . . . , p
i=1 f =1 a=1
ur , vi ≥ 0, 8r, i
wg unrestricted in sign, g = 1, . . . , h:
ð9:21Þ
This model is exactly the same as Model (9.16), the multiplier form of the
relational model, except that the multipliers corresponding to the intermediate
products are allowed to be negative. The economic interpretation of the negative
sign of the multipliers is that if Division k is producing too much of the gth
intermediate product, then its unit value wg becomes negative to discourage this
division from oversupplying this intermediate product. On the other hand, if too
much of the fth intermediate product is being produced, then wf becomes negative to
attract Division k to consume more of this intermediate product to reduce the total
value of the resources that this division consumes. An equilibrium target value for Zg
that produces the highest efficiency for the system will finally be compromised. The
constraints of the cooperative model are less stringent than that of the relational
model. The efficiency of the former is thus greater than or equal to that of the latter.
The SBM model has the same constraint set as that of the envelopment form model,
only the efficiency is calculated differently. Following Model (9.19), the slacks-
based form model of the cooperative type is:
1 m ðk Þ p
1- i=1
si- =
X
k = 1 i0
min: m
1 s p ðk Þ
1þ r=1 r
sþ = k = 1 Y r0
s
p n p
ðk Þ ðkÞ ðk Þ
s:t: λj X ij þ si- = X i0 , i = 1, . . . , m
k=1 j=1 k=1
p n p p n p
λj Z gj = λj Z gj , g = 1, . . . , h
k=1 j=1 b=1 k=1 j=1 a=1
p n p
λj Y rj - sþ
r = Y r0 , r = 1, . . . , s
k=1 j=1 k=1
ðk Þ
λj , si- , sþ
r ≥ 0, 8j, i, g, r, k:
ð9:22Þ
The models formulated in these four sections are for general systems and are thus
very complicated, but they will be much simpler for specific systems. In the next
section we use an example, which includes almost all special situations, to illustrate
the formulation of the four types of model.
9.6 An Example
Consider a DMU composed of three divisions with the structure shown in Fig. 9.3.
For Division 1, inputs X1 and X2 are supplied from outside to produce intermediate
products Z1 and Z2. A portion of Z1 is sent to Division 3 for production, and the
remainder becomes a final output, and Z2 is sent to Division 2 for production. For
Division 2, input X1 from outside and the intermediate product Z2 from Division
Fig. 9.3 Network structure Z1(1, 0)

of the example
X 1(1) , X 2(1) 1
Z1(1,3)
Z 2(1, 2)
Z 3(2,3)
X 1( 2) 2 3 Y1(3) , Y2(3)
Z3(0,3)
9.6 An Example 203
1 are consumed to produce the intermediate product Z3. For Division 3, Z1, produced
from Division 1, and Z3, a portion supplied from Division 2 and a portion supplied
from outside, are used to produce the final outputs Y1 and Y2. The four types of
models discussed in the preceding sections are formulated for this example below.
9.6.1 Independent Model
Each of the four types of models can be formulated in multiplier, envelopment, and
slacks-based forms. For the independent model the multiplier form is:
ð1Þ ð1,0Þ ð1,3Þ ð1Þ ð1,2Þ ð2Þ ð2,3Þ ð3Þ ð3Þ ð3Þ ð3Þ
max: w1 Z 10 þ Z 10 þ w2 Z 20 þ w3 Z 30 þ u1 Y 10 þ u2 Y 20
ð1 Þ ð1Þ ð1Þ ð1Þ
s:t: v1 X 10 þ v2 X 20 = 1
ð2Þ ð2Þ ð2Þ ð1,2Þ
v1 X 10 þ w2 Z 20 = 1
ð3Þ ð1,3Þ ð3Þ ð0,3Þ ð2,3Þ
w1 Z 10 þ w3 Z 30 þ Z 30 =1
ð1Þ ð1,0Þ ð1,3Þ ð1Þ ð1,2Þ ð1Þ ð1Þ ð1 Þ ð1Þ
w1 Z 1j þ Z 1j þ w2 Z 2j - v1 X 1j þ v2 X 2j ≤ 0, j = 1, . . . , n
ð2Þ ð2,3Þ ð2Þ ð2Þ ð2Þ ð1,2Þ
w3 Z 3j - v1 X 1j þ w2 Z 2j ≤ 0, j = 1, . . . , n
ð3Þ ð3Þ ð3Þ ð3Þ ð3Þ ð1,3Þ ð3Þ ð0,3Þ ð2,3Þ
u1 Y 1j þ u2 Y 2j - w1 Z 1j þ w3 Z 3j þ Z 3j ≤ 0, j = 1, . . . , n
ð3Þ ð3Þ ð1Þ ð2Þ ð1 Þ ð1Þ ð1Þ ð2Þ ð2Þ ð3Þ ð3Þ
u1 , u2 , v1 , v1 , v2 , w1 , w2 , w2 , w3 , w1 , w3 ≥0
ð9:23Þ
At optimality, the terms in square brackets in the objective function are the
efficiencies of the three divisions. The system efficiency can be defined as an
aggregation, in different forms, of the division efficiencies to be the objective
function, depending on the assumption of the study.
The envelopment form of the independent input model is:
min: θð1Þ þ θð2Þ þ θð3Þ

n
ð1Þ ð1Þ ð1Þ
s:t: λj X 1j ≤ θð1Þ X 10
j=1
n
ð1Þ ð1Þ ð1Þ
λj X 2j ≤ θð1Þ X 20
j=1
n
ð1Þ ð1,0Þ ð1,3Þ ð1,0Þ ð1,3Þ
λj Z 1j þ Z 1j ≥ Z 10 þ Z 10
j=1
n
ð1Þ ð1,2Þ ð1,2Þ
λj Z 2j ≥ Z 20
j=1
n
ð2Þ ð2Þ ð2Þ
j=1
n
λj Z 2j ≤ θð2Þ Z 20 ð9:24Þ
j=1
n
λj Z 3j ≥ Z 30
j=1
n
λj Z 1j ≤ θð3Þ Z 10
j=1
n
ð3Þ ð0,3Þ ð2,3Þ ð0,3Þ ð2,3Þ
λj Z 3j þ Z 3j ≤ θð3Þ Z 30 þ Z 30
j=1
n
ð3Þ ð3Þ ð3Þ
λj Y 1j ≥ Y 10
j=1
n
ð3Þ ð3Þ ð3Þ
λj Y 2j ≥ Y 20
j=1
ðkÞ
λj ≥ 0, 8j, k
This model is an aggregation of the three independent models, each

corresponding to one division. At optimality, θ(1), θ(2), and θ(3) are the efficiencies
of the three divisions. Again, the system efficiency must be defined by the decision
maker as an aggregation of the division efficiencies. In this case, the aggregate
efficiency serves as the objective function of the model.
The slacks-based form is:
9.6 An Example 205
1 ð1Þ - ð1Þ ð1Þ - ð1Þ

1- s1 =X 10 þ s2 =X 20
min: 2 þ
1 ð1Þþ ð1,0Þ ð1,3Þ ð1Þþ ð1,2Þ
1þs = Z 10 þ Z 10 þ s2 =Z 20
2 1
1 ð2Þ - ð2Þ ð2Þ - ð1,2Þ
1- s1 =X 10 þ s2 =Z 20
2 þ
ð2Þþ ð2,3Þ
1 þ s3 =Z 30
1 ð3Þ - ð1,3Þ ð3Þ - ð0,3Þ ð2,3Þ
1- s1 =Z 10 þ s3 = Z 30 þ Z 30
2
1 ð3Þþ ð3Þ ð3Þþ ð3Þ
1þ s1 =Y 10 þ s2 =Y 20
2
n
ð1Þ ð1Þ ð1Þ - ð1Þ
s:t: λj X 1j þ s1 = X 10
j=1
n
λj X 2j þ s2 = X 20
j=1
n
ð1Þ ð1,0Þ ð1,3Þ ð1Þþ ð1,0Þ ð1,3Þ
λj Z 1j þ Z 1j - s1 = Z 10 þ Z 10
j=1
n
ð1Þ ð1,2Þ ð1Þþ ð1,2Þ
λj Z 2j - s2 = Z 20
j=1 ð9:25Þ
n
λj X 1j þ s1 = X 10
j=1
n
ð2Þ ð1,2Þ ð2Þ - ð1,2Þ
λj Z 2j þ s2 = Z 20
j=1
n
λj Z 3j - s3 = Z 30
j=1
n
ð3Þ ð1,3Þ ð3Þ - ð1,3Þ
λj Z 1j þ s1 = Z 10
j=1
n
ð3Þ ð0,3Þ ð2,3Þ ð3Þ - ð0,3Þ ð2,3Þ
λj Z 3j þ Z 3j þ s3 = Z 30 þ Z 30
j=1
n
ð3Þ ð3Þ ð3Þþ ð3Þ
λj Y 1j - s1 = Y 10
j=1
n
λj Y 2j - s2 = Y 20
j=1
λj , s i , sðrkÞþ , sf , sðgkÞþ ≥ 0, 8j, i, r, f , g, k
The ratios in parentheses in the objective function are the efficiencies of the three
divisions.
9.6.2 Connected Model
The connected model is easier to formulate in the envelopment form, and the input-
oriented form is:
min: θ
n
ð1Þ ð1Þ ð1Þ
s:t: λj X 1j ≤ θX 10
j=1
n
ð1Þ ð1Þ ð1Þ
λj X 2j ≤ θX 20
j=1
n
ð1Þ ð1,0Þ ð1,3Þ ð1,0Þ ð1,3Þ
λj ðZ 1j þ Z 1j Þ ≥ ðZ 10 þ Z 10 Þ
j=1
n
λj Z 2j ≥ Z 20
j=1
n
ð2Þ ð2Þ ð2Þ
λj X 1j ≤ θX 10
j=1
n
λj Z 2j ≤ Z 20 ð9:26Þ
j=1
n
λj Z 3j ≥ Z 30
j=1
n
λj Z 1j ≤ Z 10
j=1
n
ð3Þ ð0,3Þ ð2,3Þ ð0,3Þ ð2,3Þ
λj ðZ 3j þ Z 3j Þ ≤ θZ 30 þ Z 30
j=1
n
ð3Þ ð3Þ ð3Þ
λj Y 1j ≥ Y 10
j=1
n
ð3Þ ð3Þ ð3Þ
λj Y 2j ≥ Y 20
j=1
ðkÞ
λj ≥ 0, 8j, k
9.6 An Example 207
Compared to the independent model (9.24), this model attaches the distance
parameter θ only to the exogenous inputs, and the distance parameter is the same
for all three divisions. In other words, it measures the system efficiency while taking
the operations of the divisions into consideration.
The merit of the envelopment form is that it shows the projection points of the
factors for the inefficient DMUs to make improvements. From the constraints of
Model (9.26), we have:
n n
ð1Þ ð1Þ ð1Þ ð1Þ ð1Þ ð1Þ
X 10 = λj X 1j , X 20 = λj X 2j ,
j=1 j=1
n n
ðoutÞ ð1Þ ð1,0Þ ð1,3Þ ð1,2ÞðoutÞ ð1Þ ð1,2Þ
Z 10 = λj Z 1j þ Z 1j , Z 20 = λj Z 2j
j=1 j=1
n n n
ð2Þ ð2Þ ð2Þ ð1,2ÞðinÞ ð2Þ ð1,2Þ ð2,3ÞðoutÞ ð2Þ ð2,3Þ
X 10 = λj X 1j , Z 20 = λj Z 2j , Z 30 = λj Z 3j
j=1 j=1 j=1
n n
ðinÞ ð3Þ ð0,3Þ ð2,3Þ ð1,3ÞðinÞ ð3Þ ð1,3Þ
Z 30 = λj Z 3j þ Z 3j , Z 10 = λj Z 1j ,
j=1 j=1
n n
Y 10 = λj Y 1j , Y 20 = λj Y 2j
j=1 j=1
ð1,0Þ ð1,3ÞðoutÞ
Note that the same output for different purposes; for example, Z 10 and Z 10
ðoutÞ ð0,3Þ ð2,3ÞðinÞ
have been combined as Z 10 , which cannot be distinguished. Z 30 and Z 30
ðinÞ
have also been combined as Z 30 and cannot be distinguished.
The multiplier form of the connected model, which is the dual of Model (9.26), is:
ð3Þ ð3Þ ð3Þ ð3Þ ð1Þ ð1,0Þ ð1,3Þ ð1Þ ð1,2Þ ð2Þ ð2,3Þ
max: u1 Y 10 þ u2 Y 20 þ w1 Z 10 þ Z 10 þ w2 Z 20 þ w3 Z 30 -
ð3Þ ð1,3Þ ð2Þ ð1,2Þ ð3Þ ð2,3Þ
w1 Z 10 þ w2 Z 20 þ w3 Z 30
ð1Þ ð1Þ ð1Þ ð1Þ ð2Þ ð2Þ ð3Þ ð0,3Þ
s:t: v1 X 10 þ v2 X 20 þ v1 X 10 þ w3 Z 30 = 1
ð1Þ ð1,0Þ ð1,3Þ ð1Þ ð1,2Þ ð1Þ ð1Þ ð1Þ ð1Þ
ð2Þ ð2,3Þ ð2Þ ð2Þ ð2Þ ð1,2Þ
w3 Z 3j - v1 X 1j þ w2 Z 2j ≤ 0, j = 1, . . . , n
ð3Þ ð3Þ ð3Þ ð3Þ ð3Þ ð1,3Þ ð3Þ ð0,3Þ ð2,3Þ
ð3Þ ð3Þ ð1Þ ð1Þ ð2Þ ð1Þ ð1Þ ð2Þ ð2Þ ð3Þ ð3Þ
u1 , u 2 , v 1 , v 2 , v 1 , w 1 , w 2 , w 2 , w 3 , w 1 , w 3 ≥ 0
ð9:27Þ
The objective value is the system efficiency, which based on the duality theorem
is equal to θ* calculated from Model (9.26). The normalization constraint corre-
sponds to the DMU, which is the sum of the three normalization constraints
corresponding to the three divisions in the independent model (9.23), with the
intermediate products moved to the objective function.
It was mentioned in the preceding section that in the connected model the system
efficiency measured from the output-oriented form under constant returns to scale
may not be the same as that measured from the input-oriented form, due to the
structure of the network. In this example, the output-oriented model is:
max: φ
n
ð1Þ ð1Þ ð1Þ
s:t: λj X 1j ≤ X 10
j=1
n
ð1Þ ð1Þ ð1Þ
λj X 2j ≤ X 20
j=1
n
ð1Þ ð1,0Þ ð1,3Þ ð1,0Þ ð1,3Þ
λj ðZ 1j þ Z 1j Þ ≥ φZ 10 þ Z 10
j=1
n
λj Z 2j ≥ Z 20
j=1
n
ð2Þ ð2Þ ð2Þ
λj X 1j ≤ X 10
j=1
n
λj Z 2j ≤ Z 20 ð9:28Þ
j=1
n
λj Z 3j ≥ Z 30
j=1
n
λj Z 1j ≤ Z 10
j=1
n
ð3Þ ð0,3Þ ð2,3Þ ð0,3Þ ð2,3Þ
λj ðZ 3j þ Z 3j Þ ≤ Z 30 þ Z 30
j=1
n
ð3Þ ð3Þ ð3Þ
λj Y 1j ≥ φY 10
j=1
n
ð3Þ ð3Þ ð3Þ
λj Y 2j ≥ φY 20
j=1
ðkÞ
λj ≥ 0, 8j, k
The production possibility set of this model is the same as that of Model (9.26).
However, the system efficiency measured from this model, which is 1/φ, is obvi-
ously not the same as that measured from Model (9.26).
9.6 An Example 209
The slacks-based form of the connected model needs some explanation. Its
constraints are:
n
λj X 1j þ s1 = X 10
j=1
n
λj X 2j þ s2 = X 20
j=1
n
ð1Þ ð1,0Þ ð1,3Þ ð1Þþ ð1,0Þ ð1,3Þ
λj Z 1j þ Z 1j - s1 = Z 10 þ Z 10
j=1
n
λj Z 2j - s2 = Z 20
j=1
n
λj X 1j þ s1 = X 10
j=1
n
ð2Þ ð1,2Þ ð2Þ - ð1,2Þ
λj Z 2j þ s2 = Z 20
j=1 ð9:29Þ
n
λj Z 3j - s3 = Z 30
j=1
n
ð3Þ ð1,3Þ ð3Þ - ð1,3Þ
λj Z 1j þ s1 = Z 10
j=1
n
ð3Þ ð0,3Þ ð2,3Þ ð3Þ - ð0,3Þ ð2,3Þ
λj Z 3j þ Z 3j þ s3 = Z 30 þ Z 30
j=1
n
λj Y 1j - s1 = Y 10
j=1
n
λj Y 2j - s2 = Y 20
j=1
λj , s i , sðrkÞþ , sf , sðgkÞþ ≥ 0, 8j, i, r, f , g, k
ð1,0Þ ð1,3Þ ð0,3Þ ð2,3Þ

where the same intermediate products Z 10 and Z 10 , and Z 30 and Z 30 have been
combined, respectively. Note that the objective function of the connected model
only involves the slacks related to the exogenous inputs and final outputs. In other
ð1,0Þ ð0,3Þ
words, only the slacks associated with Z 10 and Z 30 are needed which, however,
are not specified in Model (9.29). The objective function in this case cannot be
formulated. It can be approximated by adding two constraints of
n ð1Þ ð1,0Þ ð1Þþ ð1,0Þ n ð3Þ ð0,3Þ ð3Þ - ð0,3Þ
j = 1 λj Z 1j - s1 = Z 10 and j = 1 λj Z 3j þ s3 = Z 30 to extract out
ð1Þþ ð3Þ - ð1Þþ ð3Þ -
the required slacks of s1 and s3 from s1 and s3 , respectively. The
model is then:
ð1Þ - ð1Þ ð1Þ - ð1Þ ð2Þ - ð2Þ ð3Þ - ð0,3Þ

1- 1
4 s1 =X 10 þ s2 =X 20 þ s1 =X 10 þ s3 =Z 30
min:
ð3Þþ ð3Þ ð3Þþ ð3Þ ð1Þþ ð1,0Þ
1 þ 13 s1 =Y 10 þ s2 =Y 20 þ s1 =Z 10
ð9:30Þ
s:t: Constraint set ð9:29Þ
n
λj Z 1j - s1 = Z 10
j=1
n
ð3Þ ð0,3Þ ð3Þ - ð0,3Þ
λj Z 3j þ s3 = Z 30
j=1
ð1Þþ ð3Þ -
s1 , s3 ≥0
9.6.3 Relational Model
The relational model is formulated in multiplier form, which is:
ð3Þ ð3Þ ð1,0Þ

max: u1 Y 10 þ u2 Y 20 þ w1 Z 10
s:t: v1 X 10 þ X 10 þ v2 X 20 þ w3 Z 30 = 1
ð1,0Þ ð1,3Þ ð1,2Þ ð1Þ ð1Þ
ð2,3Þ ð2Þ ð1,2Þ
w3 Z 3j - v1 X 1j þ w2 Z 2j ≤ 0, j = 1, . . . , n
ð3Þ ð3Þ ð1,3Þ ð0,3Þ ð2,3Þ
u1 , u2 , v1 , v2 , w1 , w2 , w3 ≥ 0
ð9:31Þ
This model is the same as the connected model (9.27), in multiplier form, except
that the same factor is required to have the same multiplier. The advantage of this
model is that the system and division efficiencies can be measured at the same time.
From the constraints of this model, they are obtained as:
ð3Þ ð3Þ ð1,0Þ

E 0 = u1 Y 10 þ u2 Y 20 þ w1 Z 10
ð1Þ ð1,0Þ ð1,3Þ ð1,2Þ ð1Þ ð1Þ
E 0 = w1 Z 10 þ Z 10 þ w2 Z 20 = v1 X 10 þ v2 X 20
ð2Þ ð2,3Þ ð2Þ ð1,2Þ
E 0 = w3 Z 30 = v1 X 10 þ w2 Z 20
ð3Þ ð3Þ ð3Þ ð1,3Þ ð0,3Þ ð2,3Þ
E 0 = u1 Y 10 þ u2 Y 20 = w1 Z 10 þ w3 Z 30 þ Z 30
9.6 An Example 211
The envelopment form of this model is formulated from the dual of Model (9.31),
which is:
min: θ
n n
s:t: λj X 1j þ λj X 1j ≤ θ X 10 þ X 10
j=1 j=1
n
ð1Þ ð1Þ ð1Þ
λj X 2j ≤ θX 20
j=1
n n
ð1Þ ð1,0Þ ð1,3Þ ð1,0Þ ð3Þ ð1,3Þ
λj Z 1j þ Z 1j ≥ Z 10 þ λj Z 1j
j=1 j=1
n n
λj Z 2j ≥ λj Z 2j ð9:32Þ
j=1 j=1
n n
ð3Þ ð0,3Þ ð2,3Þ ð2Þ ð2,3Þ ð0,3Þ
λj Z 3j þ Z 3j ≤ λj Z 3j þ θZ 30
j=1 j=1
n
ð3Þ ð3Þ ð3Þ
λj Y 1j ≥ Y 10
j=1
n
ð3Þ ð3Þ ð3Þ
λj Y 2j ≥ Y 20
j=1
ðkÞ
λj ≥ 0, 8j, k
Similar to other envelopment models, this model is also able to show the
projection points of the factors for inefficient DMUs to make improvements. How-
ever, due to the combination of the same factors from different divisions, two targets
ðinÞ ðoutÞ
Z 10 and Z 30 cannot be distinguished. Based on the constraints of Model (9.32), the
targets are obtained as:
n n n
X 10 = λj X 1j þ λj X 1j , X 20 = λj X 2j
j=1 j=1 j=1
n
ðoutÞ ð1Þ ð1,0Þ ð1,3Þ
Z 10 = λj Z 1j þ Z 1j ,
j=1
n n
ðoutÞ ð1Þ ð1,2Þ ðinÞ ð2Þ ð1,2Þ
Z 20 = λj Z 2j , Z 20 = λj Z 2j
j=1 j=1
n
ðinÞ ð3Þ ð0,3Þ ð2,3Þ
Z 30 = λj Z 3j þ Z 3j ,
j=1
n n
ð3Þ ð3Þ ð3Þ ð3Þ
Y 10 = λj Y 1j , Y 20 = λj Y 2j
j=1 j=1
ðoutÞ ðinÞ
Note that Z 20 need not be the same as Z 20 .
For the slacks-based measure the constraints are essentially those describing the
production possibility set of Model (9.32). Again, since the objective function of the
slacks-based measure, as defined by Tone and Tsutsui (2009), only involves the
slacks related to the exogenous inputs and final outputs, the SBM efficiency in this
case is not defined. However, similar to the case of the connected model, it can be
approximated by separating these slacks from those combined with other slacks. The
complete model is:
1 ð1Þ - ð1Þ ð1Þ - ð1Þ ð2Þ - ð2Þ ð3Þ - ð0,3Þ

1- s1 =X 10 þ s2 =X 20 þ s1 =X 10 þ s3 =Z 30
min: 4
1 ð3Þþ ð3Þ ð3Þþ ð3Þ ð1Þþ ð1,0Þ
1þ s1 =Y 10 þ s2 =Y 20 þ s1 =Z 10
3
n n
s:t: λj X 1j þ λj X 1j ≤ X 10 þ X 10
j=1 j=1
n
λj X 1j þ s1 = X 10
j=1
n
λj X 1j þ s1 = X 10
j=1
n
λj X 2j þ s2 = X 20
j=1
n n
ð1Þ ð1,0Þ ð1,3Þ ð1,0Þ ð3Þ ð1,3Þ
λj Z 1j þ Z 1j ≥ Z 10 þ λj Z 1j
j=1 j=1
n
λj Z 1j - s1 = Z 10
j=1
n n
λj Z 2j ≥ λj Z 2j
j=1 j=1
n n
ð3Þ ð0,3Þ ð2,3Þ ð2Þ ð2,3Þ ð0,3Þ
λj Z 3j þ Z 3j ≤ λj Z 3j þ Z 30
j=1 j=1
n
ð3Þ ð0,3Þ ð3Þ - ð0,3Þ
λj Z 3j þ s3 = Z 30
j=1
n
λj Y 1j - s1 = Y 10
j=1
n
λj Y 2j - s2 = Y 20
j=1
ðk Þ ð1Þ - ð1Þ - ð2Þ - ð1Þþ ð3Þ - ð3Þþ ð3Þþ
λj , s1 , s2 , s1 , s1 , s3 , s1 , s2 ≥ 0, 8j, k
ð9:33Þ
The constraints marked with “*” are not needed to describe the production
possibility set. They are added to extract the required slacks for the objective
function to calculate the efficiency.
9.6 An Example 213
9.6.4 Cooperative Model
The cooperative model is very similar to the relational model. For the multiplier
form, it is exactly the same as Model (9.31), except that the multipliers associated
with the intermediate products are allowed to be negative:
ð3Þ ð3Þ ð1,0Þ

max: u1 Y 10 þ u2 Y 20 þ w1 Z 10
s:t: v1 X 10 þ X 10 þ v2 X 20 þ w3 Z 30 = 1
ð1,0Þ ð1,3Þ ð1,2Þ ð1Þ ð1Þ
ð2,3Þ ð2Þ ð1,2Þ
w3 Z 3j - v1 X 1j þ w2 Z 2j ≤ 0, j = 1, . . . , n
u1 , u 2 , v 1 , v 2 ≥ 0
w1 , w2 , w3 unrestricted in sign
ð9:34Þ
The efficiency of each division can be measured from the corresponding con-
straint of this model exactly the same as in the relational model case.
The envelopment form of this model is also the same as that of the relational
model, only the inequality sign of the constraints corresponding to the intermediate
products is replaced with equality sign:
min: θ
n n
s:t: λj X 1j þ λj X 1j ≤ θ X 10 þ X 10
j=1 j=1
n
ð1Þ ð1Þ ð1Þ
λj X 2j ≤ θX 20
j=1
n n
ð1Þ ð1,0Þ ð1,3Þ ð1,0Þ ð3Þ ð1,3Þ
λj Z 1j þ Z 1j = Z 10 þ λj Z 1j
j=1 j=1
n n
λj Z 2j = λj Z 2j ð9:35Þ
j=1 j=1
n n
ð3Þ ð0,3Þ ð2,3Þ ð2Þ ð2,3Þ ð0,3Þ
λj Z 3j þ Z 3j = λj Z 3j þ θZ 30
j=1 j=1
n
ð3Þ ð3Þ ð3Þ
λj Y 1j ≥ Y 10
j=1
n
ð3Þ ð3Þ ð3Þ
λj Y 2j ≥ Y 20
j=1
ðkÞ
λj ≥ 0, 8j, k
At optimality, the target value of the input, output, and intermediate product can
be calculated similar to the relational model case.
The system efficiency of the slacks-based form, similar to the relational model
case, cannot be defined because some slacks are confounded with others. Following
the idea introduced in the relational model, the system efficiency in this case can be
approximated via the following model:
1 ð1Þ - ð1Þ ð1Þ - ð1Þ ð2Þ - ð2Þ ð3Þ - ð0,3Þ

1- s1 =X 10 Þþ; ðs2 =X 20 Þ; þ; ðs1 =X 10 Þ; þ; ð
s3 =Z 30
4
min:
1 ð3Þþ ð3Þ ð3Þþ ð3Þ ð1Þþ ð1,0Þ
1þ s1 =Y 10 Þ þ ðs2 =Y 20 Þ þ ð
s1 =Z 10
3
n n
s:t: λj X 1j þ λj X 1j ≤ X 10 þ X 10
j=1 j=1
n
λj X 1j þ s1 = X 10
j=1
n
λj X 1j þ s1 = X 10
j=1
n
λj X 2j þ s2 = X 20
j=1
n n
ð1Þ ð1,0Þ ð1,3Þ ð1,0Þ ð3Þ ð1,3Þ
λj ðZ 1j þ Z 1j Þ = Z 10 þ λj Z 1j
j=1 j=1
n
λj Z 1j - s1 = Z 10
j=1
n n
λj Z 2j = λj Z 2j
j=1 j=1
n n
ð3Þ ð0,3Þ ð2,3Þ ð2Þ ð2,3Þ ð0,3Þ
λj ðZ 3j þ Z 3j Þ = λj Z 3j þ Z 30
j=1 j=1
n
ð3Þ ð0,3Þ ð3Þ - ð0,3Þ
λj Z 3j þ s3 = Z 30
j=1
n
λj Y 1j - s1 = Y 10
j=1
n
λj Y 2j - s2 = Y 20
j=1
ðkÞ ð1Þ - ð1Þ - ð2Þ - ð1Þþ ð3Þ - ð3Þþ ð3Þþ
λj , s1 , s2 , s1 , s1 , s3 , s1 , s2 ≥ 0, 8j, k
ð9:36Þ
This chapter introduces four basic ideas for modeling network systems. In real-
world applications different assumptions have been imposed on measuring the
efficiency of a system. Most of them are variations of these four basic models, and
they will be discussed in detail in the related chapters.
There are many studies on applications of the independent model for network
systems, with Wang et al. (1997), Seiford and Zhu (1999), and Sexton and Lewis
(2003) being the most famous ones. The first study using a network approach to
discuss the efficiency measurement for network systems is probably Färe and
Whittaker (1995), and their model is the connected one. Other studies from Färe
and his colleagues related to the modeling of network systems based on the
connected approach include Färe and Grosskopf (1996a, 1996b) and Färe et al.
(2007, 2014). Avkiran and Goto (2011), Fukuyama and Mirdehghan (2012), von
Geymueller (2009), An et al. (2015), and Yu (2010) have used connected models,
with different objective functions, to identify non-dominated DMUs. The relational
model is attributed to Kao (2009), while Kao (2013, 2014a) are later contributions to
this. The slacks-based measure of efficiency for network systems has attracted many
studies in recent years, such as Fukuyama and Weber (2010), Avkiran and
McCrystal (2012), Akther et al. (2013), Kao (2014b), Tone and Tsutsui (2014a,
2014b), Lozano and Gutiérrez (2014), Moreno and Lozano (2014), Lozano (2015),
and Lozano (2016). The idea of the cooperative model is the same as that of
continuity discussed in Tone and Tsutsui (2009) which has been studied in Chen
et al. (2016), Cui and Li (2016), Lu et al. (2016), Avkiran and McCrystal (2012),
Chang et al. (2017), Lozano (2017), Li and Cui (2017), Li et al. (2015), Tavana et al.
(2013), Lin and Chiu (2013), Sánchez-González et al. (2017), Xu and Cui (2017),
and Tone and Tsutsui (2010) for different structures of network systems, with
various applications.
Different data types are another stream of study in network data envelopment
analysis. For fuzzy data, we have Kao (2014d), in that a property that the system
efficiency slack is the sum of the division efficiency slacks, which holds for
deterministic cases, also holds here. Lozano and Moreno (2014) extended several
fuzzy approaches for whole-unit systems to network ones. For stochastic data, Kao
and Liu (2022) developed a model to calculate efficiencies with correlated stochastic
data for network systems.
Network DEA models have also been classified or decomposed differently in
other studies, see, for example, Castelli et al. (2010), Cook et al. (2010a), Cook et al.
(2010b), Kao and Hwang (2010), Chen et al. (2013), and Kao (2014c).
References
Akther, S., Fukuyama, H., & Weber, W. L. (2013). Estimating two-stage network slacks-based
inefficiency: An application to Bangladesh banking. Omega, 41, 88–96.
An, Q. X., Chen, H. X., Wu, J., & Liang, L. (2015). Measuring slacks-based efficiency for
commercial banks in China by using a two-stage DEA model with undesirable output. Annals
of Operations Research, 235, 13–35.
Avkiran, N. K., & Goto, M. (2011). A tool for scrutinizing bank bailouts based on multi-period peer
benchmarking. Pacific-Basin Finance Journal, 19, 447–469.
Avkiran, N. K., & McCrystal, A. (2012). Sensitivity analysis of network DEA: NSBM versus
NRAM. Applied Mathematics and Computation, 218, 11226–11239.
Castelli, L., Pesenti, R., & Ukovich, W. (2010). A classification of DEA models when the internal
structure of the decision making units is considered. Annals of Operations Research, 173,
207–235.
Chang, Y. T., Lee, S., & Park, H. (2017). Efficiency analysis of major cruise lines. Tourism
Management, 58, 181–186.
Chen, Y., Cook, W. D., Kao, C., & Zhu, J. (2013). Network DEA pitfalls: Divisional efficiency and
frontier projection under general network structures. European Journal of Operational
Research, 226, 507–515.
Chen, Y., Li, Y. J., Liang, L., Salo, A., & Wu, H. Q. (2016). Frontier projection and efficiency
decomposition in two-stage processes with slacks-based measures. European Journal of Oper-
ational Research, 250, 543–554.
Cui, Q., & Li, Y. (2016). Airline energy efficiency measures considering carbon abatement: A new
strategic framework. Transportation Research Part D-Transport and Environment, 49,
246–258.
Cook, W. D., Zhu, J., Bi, G. B., & Yang, F. (2010a). Network DEA: Additive efficiency
decomposition. European Journal of Operational Research, 207, 1122–1129.
Cook, W. D., Liang, L., & Zhu, J. (2010b). Measuring performance of two-stage network structures
by DEA: A review and future perspective. Omega, 38, 423–430.
Färe, R., & Grosskopf, S. (1996a). Intertemporal production Frontiers: With dynamic DEA.
Kluwer Academic.
Färe, R., & Grosskopf, S. (1996b). Productivity and intermediate products: A frontier approach.
Economics Letters, 50, 65–70.
Färe, R., Grosskopf, S., & Whittaker, G. (2007). Network DEA. In J. Zhu & W. D. Cook (Eds.),
Modeling data irregularities and structural complexities in data envelopment analysis
Färe, R., Grosskopf, S., & Whittaker, G. (2014). Network DEA II. In W. D. Cook & J. Zhu (Eds.),
Data envelopment analysis: A handbook on the modeling of internal structures and networks
Färe, R., & Whittaker, G. (1995). An intermediate input model of dairy production using complex
survey data. Journal of Agricultural Economics, 46, 201–213.
Fukuyama, H., & Mirdehghan, S. M. (2012). Identifying the efficiency status in network DEA.
Fukuyama, H., & Weber, W. L. (2010). A slacks-based inefficiency measure for a two-stage system
with bad outputs. Omega, 38, 398–409.
Kao, C. (2009). Efficiency decomposition in network data envelopment analysis: A relational
model. European Journal of Operational Research, 192, 949–962.
Kao, C. (2013). Dynamic data envelopment analysis: A relational analysis. European Journal of
References 217
Kao, C. (2014a). Efficiency decomposition in network data envelopment analysis. In W. D. Cook &
J. Zhu (Eds.), Data envelopment analysis: A handbook on the modeling of internal structures
and networks (pp. 55–77). Springer.
Kao, C. (2014b). Efficiency decomposition in network data envelopment analysis with slacks-based
measure. Omega, 45, 1–6.
Kao, C. (2014c). Network data envelopment analysis: A review. European Journal of Operational
Kao, C. (2014d). Network data envelopment analysis with fuzzy data. In A. Emrouznejad &
M. Tavana (Eds.), Performance measurement with fuzzy data envelopment analysis
Kao, C. (2018). A classification of slacks-based efficiency measures in network data envelopment
analysis with an analysis of the properties possessed. European Journal of Operational
Research 270, 1109–1121.
Kao, C., & Liu, S. T. (2022). Stochastic efficiencies of network production systems with correlated
stochastic data: The case of Taiwanese commercial banks. Annals of Operations Research, 315,
1151–1174.
Li, Y., & Cui, Q. (2017). Carbon neutral growth from 2020 strategy and airline environmental
inefficiency: A network range adjusted environmental data envelopment analysis. Applied
Energy, 199, 13–24.
Li, Y., Wang, Y. Z., & Cui, Q. (2015). Evaluating airline efficiency: An application of virtual
frontier network SBM. Transportation Research Part E-Logistics and Transportation Review,
81, 1–17.
Lin, T. Y., & Chiu, S. H. (2013). Using independent component analysis and network DEA to
improve bank performance evaluation. Economic Modelling, 32, 608–616.
Lozano, S. (2015). Alternative SBM model for network DEA. Computers & Industrial Engineer-
ing, 82, 33–40.
Lozano, S. (2016). Slacks-based inefficiency approach for general networks with bad outputs: An
application to the banking sector. Omega, 60, 73–84.
Lozano, S. (2017). Technical and environmental efficiency of a two-stage production and abate-
ment system. Annals of Operations Research, 255, 199–219.
Lozano, S., & Gutiérrez, E. (2014). A slacks-based network DEA efficiency analysis of European
airlines. Transportation Planning and Technology, 37, 623–637.
Lozano, S., & Moreno, P. (2014). Network fuzzy data envelopment analysis. In A. Emrouznejad &
M. Tavana (Eds.), Performance measurement with fuzzy data envelopment analysis
Lu, W. M., Kweh, Q. L., Nourani, M., & Huang, F. W. (2016). Evaluating the efficiency of dual-use
technology development programs from the R&D and socio-economic perspectives. Omega,
62, 82–92.
Moreno, P., & Lozano, S. (2014). A network DEA assessment of team efficiency in the NBA.
Sánchez-González, C., Sarto, J. L., & Vicente, L. (2017). The efficiency of mutual fund companies:
Evidence from an innovative network SBM approach. Omega, 71, 114–128.
Seiford, L. M., & Zhu, J. (1999). Profitability and marketability of the top 55 US commercial banks.
Sexton, T. R., & Lewis, H. F. (2003). Two-stage DEA: An application to Major League Baseball.
Tavana, M., Mirzagoltabar, H., Mirhedayatian, S. M., & Saen, R. F. (2013). A new network
epsilon-based DEA model for supply chain performance evaluation. Computer & Industrial
Engineering, 66, 501–513.
Tone, K., & Tsutsui, M. (2009). Network DEA: A slacks-based measure approach. European
Tone, K., & Tsutsui, M. (2010). Dynamic DEA: A slacks-based measure approach. Omega, 38,
145–156.
Tone, K., & Tsutsui, M. (2014a). Dynamic DEA with network structure: A slacks-based measure
approach. Omega, 42, 124–131.
Tone, K., & Tsutsui, M. (2014b). Slacks-based network DEA. In W. D. Cook & J. Zhu (Eds.), Data
envelopment analysis: A handbook on the modeling of internal structures and networks
von Geymueller, P. (2009). Static versus dynamic DEA in electricity regulation: The case of US
transmission system operators. Central European Journal of Operations Research, 17,
397–413.
Wang, C. H., Gopal, S., & Zionts, S. (1997). Use of data envelopment analysis in assessing
information technology impact on firm performance. Annals of Operations Research, 73,
191–213.
Xu, X., & Cui, Q. (2017). Evaluating airline energy efficiency: An integrated approach with
network epsilon-based measure and network slacks-based measure. Energy, 122, 274–286.
Yu, M. M. (2010). Assessment of airport performance using the SBM-NDEA model. Omega, 38,
440–452.
Chapter 10
Basic Two-Stage Systems
The simplest network structure is that of two divisions, in that the operations of the
system are divided into two parts, performed by the two divisions. This type of
system was first noticed by Charnes et al. (1986) in studying the performance of US
Army recruitment. They found that army recruitment actually had two stages:
creating awareness through advertising and signing contracts. To assess the impact
of an input on the performance of the system, it is necessary to know the division that
this input is associated with, so that the true effect of this input can be identified.
Two-stage systems started attracting the attention of researchers when it was
noted that the effects of some inputs on the outputs can be indirect. For example, it is
expected that IT (information technology) should have a positive impact on the
performance of banks. However, Cron and Sobol (1983) found that this may not
always be true. A later study (Wang et al. 1997) found that the operation of banking
and similar industries had two stages: capital collection and investment. The contri-
bution of IT was to the former, and whether a bank would actually make a profit or
not was dependent on correct investment decisions being made. Many problems
have this two-stage structure, and this chapter concentrates on a special two-stage
system, the basic two-stage one, where all the inputs supplied from outside are
consumed by the first division to produce the intermediate products for the second
division to produce the final outputs. The first division does not produce final
outputs, and the second division does not consume exogenous inputs.
The DEA methodology applied to the basic two-stage system is called the
two-stage DEA. However, the two-stage DEA to be discussed in this chapter should
be distinguished from another one, for which the studies first apply a DEA model to
measure the efficiency of a set of DMUs, and then in the second stage apply a
regression analysis to find the relationship between the efficiency measures and
some exogenous factors. The discussion in this chapter is limited to those models
that have been used to measure the efficiency of the basic two-stage system, which
are classified as the independent model, ratio-form measures, distance function
measures, and slacks-based measures.
https://doi.org/10.1007/978-3-031-27593-7_10
220 10 Basic Two-Stage Systems
10.1 Independent Model
In the basic two-stage system only one division has exogenous inputs and one has
final outputs, the superscript used to distinguish the divisions they are associated
with, as introduced in Chap. 9, thus becomes unnecessary. Figure 10.1 shows the
structure of the basic two-stage system, where the first division consumes all the
exogenous inputs Xi, i = 1, . . ., m, to produce the intermediate products Zg, g = 1,
. . ., h, which in turn all are consumed by the second division to produce the final
outputs Yr, r = 1,. . ., s. The most straightforward way to investigate the performance
of this type of system is to use the independent model that treats the two divisions as
independent DMUs to measure their efficiencies. The whole system can also be
treated as a whole unit, or a so-called black box, and its efficiency can be measured
by applying the conventional DEA models. In this chapter the terms division and
stage will be used interchangeably.
Consider a simple case of five DMUs, using one-input X to produce one inter-
mediate product Z in the first stage and using the intermediate product Z to produce
one output Y in the second stage, with the data shown in columns two to four of
Table 10.1. Figure 10.2 shows the production frontiers of the two stages, where the
right part shows the five DMUs, superscripted by (1), using X to produce Z, and the
left part shows the five DMUs, superscripted by (2), using Z to produce Y. Note that
the second stage has been rotated counterclockwise by 90 degrees so that the two
stages can use the same coordinate for the intermediate product Z. When the first
stage is considered as an independent production process, the frontier constructed
from the five DMUs under constant returns to scale is ray OC(1).
Column five of Table 10.1 shows the efficiencies of the five DMUs, in that only
DMU C is efficient. Similarly, when the second stage is considered as an indepen-
dent production process, then OD(2) will be the frontier, with the efficiencies of the
Fig. 10.1 Structure of the System

basic two-stage system
Xi Zg Yr
1 2
i=1,…, m g =1,…, h r =1,…, s
Table 10.1 Data and independent efficiency measures for the example
DMU X Z Y E0 E0 E0 (Combined) E0 E0 E0 (Combined)
A 2 1 0.5 1/2 2/5 1/4 (1/5) 1 1 1 (1)
B 4 2 1 1/2 2/5 1/4 (1/5) 5/8 1/2 2/7 (5/16)
C 3 3 2 1 8/15 2/3 (8/15) 1 4/7 1 (4/7)
D 5 4 5 4/5 1 1 (4/5) 9/10 1 1 (9/10)
E 6 5 5.5 5/6 22/25 11/12 (11/15) 1 1 1 (1)
Stage 2 Stage 1
6
E (2)
E (1)
D ( 2) 4 D (1)
C ( 2)
C (1)
B ( 2) B
( 2)
2 B̂ (1)
B̂ ( 2) B (1)
B (1)
A (2)
A(1)
Y 6 4 2 O 2 4 6 X
Fig. 10.2 Frontiers of the two stages of production of the example
Fig. 10.3 Frontiers of the

black-box and combined Y
technologies of the example
8
6 Ê
E=E
D
D̂ D
4
Ĉ
C
B̂
BS
2 B C
Â
BO
BT
B
A=A
O 2 4 6 X
five DMUs shown in column six, in that only DMU D is efficient. None of the
DMUs is efficient in both stages. If the whole system is treated as a black box where
input X is used to produce output Y, then DMU D is efficient. Figure 10.3 shows the
production of the five DMUs, where ray OD is the frontier. The efficiencies
measured from the CCR model are shown in column seven.
The CCR model applied to measure the efficiency of the system directly does not
consider the operations of individual stages. To see the effect, consider, for example,
DMU B. This DMU uses 4 units of X to produce 2 units of Z in the first stage, with
ð1Þ
the projection point of B X ð1Þ = 2 from the input side. In the second stage this
B
ð2Þ
DMU uses 2 units of Z to produce 1 unit of Y, with the projection point of B
Y ð 2Þ = 2:5 from the output side. Combining these two stages together, this DMU
B
should be able to use X ð 1Þ = 2 units of X to produce Y ð2Þ = 2:5 units of Y, if it is
B B
efficient in both stages, with a rate of Y ð 2Þ =X ð 1Þ = 2:5=2: In reality, DMU B uses
B B
X Bð1Þ = 4 units of X to produce Y Bð2Þ = 1 unit of Y, with a rate of Y Bð2Þ =X Bð1Þ = 1=4:
The combined efficiency of DMU B is thus (1/4)/(2.5/2) = 1/5. This idea can be
applied to measure the combined efficiencies for other DMUs. The results are shown
in parentheses in column eight of Table 10.1, where all values are smaller than the
corresponding ones measured from the black-box model. Ray OA in Fig. 10.3 is the
frontier corresponding to the combined technology, which obviously has a larger
slope than that corresponding to the black-box technology, ray OD. None of the five
DMUs is efficient under the combined technology.
If the two stages are operating under variable returns to scale, then the
corresponding frontiers are the connected line segments A(1)C(1)E(1) for stage one
and A(2)D(2)E(2) for stage two. The fourth- to second-to-last columns of Table 10.1
show the efficiencies of the five DMUs for stage one, stage two, and the system
measured separately via the BCC model. In stage one, DMUs A, C, and E are
efficient, and in stage two DMUs A, D, and E are efficient. For the black-box system
the frontier is the connected line segments ADE, where only DMU B is inefficient, as
shown in Fig. 10.3.
The system efficiency under variable returns to scale can also be measured via the
two efficiency frontiers, taking the operations of the two stages into account. Using
DMU B to explain this, it uses 4 units of X to produce 2 units of Z in the first stage.
The frontier from the input side, however, indicates that only X Bð1Þ = 2.5 units are
needed to produce the same amount of Z. In the second stage, this DMU uses 2 units
of Z to produce 1 unit of Y, where an amount of Y Bð2Þ = 2 units is expected if it
operates efficiently. Altogether, this DMU uses 4 units of X to produce 1 unit of Y,
with a rate of 1/4, while an amount of 2.5 units of X is able to produce 2 units of Y,
with a rate of 2/2.5, if it is efficient in both stages. The combined efficiency is thus
(1/4)/(2/2.5) = 5/16. The values in parentheses in the last column of Table 10.1 are
the combined efficiencies of the five DMUs measured under this combined technol-
ogy. Similar to the constant returns to scale case, all values are less than or equal to
the corresponding ones measured from the black-box model. The connected line
segments ACDE show the corresponding frontier, which lies above that
corresponding to the black-box technology, ADE.
10.2 Ratio-Form Efficiency Measures 223
This example shows that each division has its own technology for production, and
the combined technology of the two divisions may not be the same as the black-box
technology obtained without considering the intermediate product. In measuring the
system efficiency the operations of the two divisions must thus be taken into account
to be able to identify inefficient DMUs.
10.2 Ratio-Form Efficiency Measures
As discussed in Chap. 2, one branch of efficiency measurement is to express

efficiency as a ratio of the aggregate output to aggregate input. This branch of
models can be classified into two types: efficiency decomposition and efficiency
aggregation. For efficiency decomposition the system efficiency is expressed as the
ratio of the aggregate exogenous output to the aggregate exogenous input, excluding
the intermediate products. Due to the structure of the model, the system efficiency
can be decomposed into the division efficiencies in specific mathematical forms. The
efficiency aggregation, in contrast, defines the system efficiency in a pre-specified
mathematical form of the division efficiencies, where the intermediate products are
included. Several models have been proposed under these two ideas.
10.2.1 Efficiency Decomposition
In order to reflect the true performance of a basic two-stage system, the operations of
the two divisions must be taken into account in measuring the efficiency of a DMU.
The operations of the divisions, as expressed in Kao and Hwang (2008), require the
aggregate output to be less than or equal to the aggregate input, and the same
multiplier is applied to the same factor in the aggregation. In other words, the system
efficiency is measured under the technologies of both the system and the two
divisions.
10.2.1.1 Constant Returns to Scale
Based on the structure of the basic two-stage system shown in Fig. 10.1, the model
proposed by Kao and Hwang (2008) under constant returns to scale in multiplier
form is:
s
E0 = max: ur Y r0
r=1
m
s:t: vi X i0 = 1
i=1
System constraints :
s m
ur Y rj - vi X ij ≤ 0, j = 1, . . . , n
r=1 i=1 ð10:1Þ
Division constraints :
h m
wg Z gj - vi X ij ≤ 0, j = 1, . . . , n
g=1 i=1
s h
ur Y rj - wg Z gj ≤ 0, j = 1, . . . , n
r=1 g=1
ur , vi , wg ≥ 0, r = 1, . . . , s, i = 1, . . . , m, g = 1, . . . , h
Note that the non-Archimedean number ε that should be imposed on the multi-
pliers to avoid the corresponding factor being ignored has been omitted to make the
expression simpler. The characteristic of this model is that the multiplier wg associ-
ated with the intermediate product Zg as the output of division one is required to be
the same as that the input of division two. According to the classification of Chap. 9,
this is a relational model.
ðk Þ
At optimality, the system efficiency E0 and the division efficiencies E0 , based on
the constraints of Model (10.1), can be expressed as:
s m
E0 = ur Y r0 = vi X i0
r=1 i=1
h m
ð1Þ
E0 = wg Z g0 = vi X i0 ð10:2Þ
g=1 i=1
s h
ð2Þ
E0 = ur Y r0 = wg Z g0
r=1 g=1
Notably, the system efficiency is the product of the two-division efficiencies:
h s s
ð1Þ ð2Þ g = 1 wg Z g0 r = 1 ur Y r0 r = 1 ur Y r0
E0 × E0 = m × h
= m = E0

i = 1 vi X i0 g = 1 wg Z g0 i = 1 vi X i0
The characteristic of the mathematical relationship of the product in this case is

that a system with two divisions of similar efficiencies is better than those with one
high and one low efficiency, because the former will have a larger product. And this
provides a general guideline for a DMU to have higher system efficiencies.
Table 10.2 Decomposition of the efficiency measures of the example

Division 1 Division 2 System
DMU E0 T0 S0 E0 T0 S0 E0 T0 S0
A 1/2 1 1/2 2/5 1 2/5 1/5 1 1/5
B 1/2 5/8 4/5 2/5 1/2 4/5 1/5 5/16 16/25
C 1 1 1 8/15 4/7 14/15 8/15 4/7 14/15
D 4/5 9/10 8/9 1 1 1 4/5 9/10 8/9
E 5/6 1 5/6 22/25 1 22/25 11/15 1 11/15
A closer examination of the constraints corresponding to the system and divisions

in Model (10.1) reveals that the sum of the two-division constraints corresponding to
a DMU is equal to the system constraint corresponding to the same DMU. The
system constraints are thus redundant and can be deleted to make the model simpler.
To further explain the product relationship, consider the same example discussed
in Sect. 10.1. By applying Model (10.1) to the data in Table 10.1, the optimal
solution obtained for DMU B is (u, v, w) = (0.2, 0.25, 0.25), with an objective
value of 0.2, which is also the system efficiency. Based on Expression (10.2), the
division efficiencies for this DMU are 0.5 and 0.4. The efficiencies of the other four
DMUs can be calculated similarly, with the results shown in columns two, five, and
eight of Table 10.2. The product of the two-division efficiencies is clearly equal to
the system efficiency. It is also noted that, in this special case, the division efficien-
cies measured from Model (10.1) are the same as those measured from the indepen-
dent model, shown in columns five and six of Table 10.1.
Since there may exist multiple solutions for Model (10.1), the results of
decomposing the system efficiency into the two-division efficiencies may not be
unique, which makes the efficiencies of different DMUs incomparable. To address
this shortcoming, Kao and Hwang (2008) proposed a two-step approach, in that the
system efficiency E0 is measured via Model (10.1) in the first step, and the efficiency
of the division which is considered more important is measured by fixing the system
efficiency at E0 in the second step. Suppose the first division is considered more
ð1Þ
important. The efficiency of the first division E 0 is measured via the following
model:
h
ð1Þ
E 0 = max: wg Z g0
g=1
m
s:t: vi X i0 = 1
i=1
s m
ur Y r0 = E 0 vi X i0
r=1 i=1 ð10:3Þ
h m
wg Z gj - vi X ij ≤ 0, j = 1, . . . , n
g=1 i=1
s h
ur Y rj - wg Z gj ≤ 0, j = 1, . . . , n
r=1 g=1
ur , vi , wg ≥ 0, r = 1, . . . , s, i = 1, . . . , m, g = 1, . . . , h:
ð1Þ
After E 0 is solved from this model, the efficiency of the second division is
ð2Þ ð1Þ
calculated as E0 = E 0 =E 0 . This two-step approach can be combined into one step,
as described in Liu (2011), to simplify the work. Liang et al. (2008) showed that the
division efficiencies measured from Model (10.1) will be the same as those measured
from the independent model when there is only one intermediate product. In this case
the efficiency decomposition will be unique.
Model (10.1) is formulated from the input side. The relational model can also be
formulated from the output side, in which case it is “min. m i = 1 vi X i0 /
s
r = 1 ur Y r0
m s
subject to the same constraints.” Since minimizing v X
i = 1 i i0 / r = 1 ur Y r0 is
equivalent to maximizing sr = 1 ur Y r0 / m v X
i = 1 i i0 , the system efficiency measured
from the output model is the same as that measured from the input one, and they
yield the same decomposition for the division efficiencies.
10.2.1.2 Variable Returns to Scale
In the conventional black-box model the difference between the efficiencies under
constant and variable returns to scale is due to economies of scale. Specifically, the
ratio of the CCR efficiency to the BCC efficiency is the scale efficiency. This type of
efficiency decomposition is also applicable to the basic two-stage system, as
discussed in Kao and Hwang (2011, 2014). The idea is to keep the amounts of the
intermediate products fixed and use an input model to find the extent that the inputs
can be reduced in the first stage, and use an output model to find the extent that the
outputs can be expanded in the second stage. Keeping the intermediate products
fixed avoids the possibility of yielding conflict targets for the intermediate products
as the output of stage 1 and the input of stage 2.
To measure the technical and scale efficiencies, we first measure the overall
efficiency via Model (10.1), and Model (10.3) is applied if necessary. For stage
ð1Þ
1, the input technical efficiency T 0 is measured via the following model by fixing
ð1Þ
the overall efficiency of this stage at E 0 :
h
ð1Þ
T 0 = max: ~ g Z g0 - w
w ~0
g=1
m
s:t: ~vi X i0 = 1
i=1
Constant returns to scale constraints :
h m
ð1Þ
wg Z g0 = E 0 vi X i0
g=1 i=1
h m
wg Z gj - vi X ij ≤ 0, j = 1, . . . , n
g=1 i=1
s h
ur Y rj - wg Z gj ≤ 0, j = 1, . . . , n
r=1 g=1
Variable returns to scale constraints :
h m
w ~0 -
~ g Z gj - w ~vi X ij ≤ 0, j = 1, . . . , n
g=1 i=1
ur , vi , ~vi , wg , w
~ g ≥ 0, r = 1, . . . , s, i = 1, . . . , m, g = 1, . . . , h:
ð10:4Þ
ð1Þ
After the technical efficiency T 0 is calculated, the scale efficiency is calculated
ð1Þ ð1Þ ð1Þ
as S0 =E 0 /T 0 .
For stage 2, the technical efficiency is measured from the output side. The idea is
the same as that of stage 1, except the maximization of the input technical efficiency
is changed to the minimization of the inverse of the output technical efficiency. The
model is:
h
ð2Þ
1=T 0 = min: ~ g Z g0 þ w
w ~0
g=1
s
s:t: ~ur Y r0 = 1
r=1
Constant returns to scale constraints :
s h
ð2Þ
ur Y r0 = E 0 wg Z g0
r=1 g=1
h m ð10:5Þ
wg Z gj - vi X ij ≤ 0, j = 1, . . . , n
g=1 i=1
s h
ur Y rj - wg Z gj ≤ 0, j = 1, . . . , n
r=1 g=1
Variable returns to scale constraints :
s h
~ur Y rj - ~ g Z gj þ w
w ~ 0 ≤ 0, j = 1, . . . , n
r=1 g=1
ur , ~ur , vi , wg , w
~ g ≥ 0, r = 1, . . . , s, i = 1, . . . , m, g = 1, . . . , h:
ð2Þ
After the technical efficiency T 0 is calculated from this model, the scale
ð2Þ ð2Þ ð2Þ
efficiency is calculated as S0 =E0 /T 0 .
The technical and scale efficiencies of the DMU are calculated as the product of
those of the two stages, respectively. We thus have the following decompositions:

E 0 = E 0 × E 0 = T 0 × S0 × T 0 × S0 ð10:6aÞ

E 0 = T 0 × S0 = T 0 × T 0 × S0 × S 0 ð10:6bÞ
where T0 and S0 are the technical and scale efficiencies of the DMU, respectively.
Referring to Fig. 10.2, the production frontiers under the variable returns to scale
technology for stages 1 and 2 are the connected line segments A(1)C(1)E(1) and
A(2)D(2)E(2), respectively. DMU B, for example, applies X Bð1Þ = 4 units of X to
produce 2 units of Z in the first stage. From the input side this stage only needs
X Bð1Þ = 2:5 units of X to produce the same amount of Z, provided it is efficient. The
input technical efficiency is X Bð1Þ =X Bð1Þ = 5=8: In the second stage this DMU applies
2 units of Z to produce Y Bð2Þ = 1 unit of Y. From the output side, if this DMU is
efficient, then it will be able to produce Y Bð2Þ = 2 units of Y using the current amount
of the intermediate input. The output technical efficiency is Y Bð2Þ = Y Bð2Þ = 1=2: The
combined technical efficiency of this DMU, as discussed in Sect. 10.1, is 5/16,
which is the product of the technical efficiencies of the two divisions,
(5/8) × (1/2) = 5/16.
The discussion of the scale efficiency is similar. The production frontiers for the
two stages under constant returns to scale are rays OC(1) and OD(2). The input scale
efficiency of stage 1 is X ð1Þ =X Bð1Þ = 4=5, which is the ratio of the input overall
B
efficiency to the input technical efficiency, X ð 1Þ =X Bð1Þ = X Bð1Þ =X Bð1Þ , and the
B
output scale efficiency of stage 2 is Y Bð2Þ =Y ð 2Þ = 4=5, which is the ratio of the output
B
overall efficiency to the output technical efficiency, Y Bð2Þ =Y ð 2Þ = Y Bð2Þ =Y Bð2Þ :

B
From the perspective of the system, if DMU B is technically efficient in both stages,
then it is able to use 2.5 units of X to produce 2 units of Y via Z = 2, with a rate of
2/2.5. If it is overall efficient in both stages, then it only needs 2 units of X to produce
2.5 units of Y, with a rate of 2.5/2. The combined scale efficiency of the two stages is
(2/2.5)/(2.5/2), which is just the product of the input scale efficiency of stage 1 and
output scale efficiency of stage 2, (4/5) × (4/5) = 16/25. Finally, the product of the
combined technical and combined scale efficiencies is just the combined overall
efficiency, (5/16) × (16/25) = 1/5. Table 10.2 shows the overall, technical, and scale
efficiencies of the system and the two divisions. The results clearly satisfy the
relationships stated in Expressions (10.6a) and (10.6b).
This example shows that the system efficiency can be decomposed into the
product of the two-division efficiencies from an overall perspective, and the overall
efficiency of the first stage can be further decomposed into the product of the input
technical and input scale efficiencies, and that of the second stage can be further
decomposed into the product of the output technical and output scale efficiencies.
Moreover, the technical efficiency of the system is the product of those of the two
stages, as is the scale efficiency of the system. For both the system and the two
stages, the overall efficiency is the product of the technical and scale efficiencies.
In Fig. 10.3 the frontiers of the combined technologies under constant and
variable returns to scale, i.e., ray OA and the connected line segments ACDE, do
not overlap, indicating that no DMU is scale efficient. This is because none of the
five DMUs is scale efficient in both stages. Using DMU B to explain the overall,
technical, and scale efficiencies, it uses 4 units of X to produce 1 unit of Y (point B).
If it is efficient under constant returns to scale, then it can reduce the input needed
from 4 units to 2 units (point BO) and increase the output from 1 unit to 2.5 units
(point B). This is a situation of using 4 units of X to produce 1 unit of Y compared to
using 2 units of X to produce 2.5 units of Y. The overall efficiency is thus (1/4)/(2.5/
2) = 1/5. If this DMU is efficient under variable returns to scale, then it can reduce
the input to 2.5 units (point BT) and increase the outputs to 2 units (point B), with a
technical efficiency of (1/4)/(2/2.5) = 5/16. The difference between technically
efficient and overall efficient is using 2.5 units of X to produce 2 units of Y (point
B), versus using 2 units of X (point BS) to produce 2.5 units of Y (point B). The scale
efficiency is thus (2/2.5)/(2.5/2) = 16/25. The overall efficiency, 1/5, is again the
product of the technical and scale efficiencies, (5/16) × (16/25).
10.2.1.3 Game-Theoretic Approach
The two divisions in the basic two-stage system can be considered as two players in a
nonzero-sum game, with efficiency measures as the payoffs to gain.
Let b(k) be the breakdown point of player k, k = 1, 2, and E(k)(u, v, w) be the
efficiency that player k will obtain if the strategy (u, v, w) is applied. The breakdown
point represents the possible payoff that one player is able to obtain without
bargaining with the other player. Its choice is based on modeling judgment. Nash
(1950, 1953) showed that if the feasible set associated with this game is compact,
convex, and contains some payoffs such that each player’s payoff is at least as large
as its breakdown payoff, then a unique solution can be obtained by solving the
problem of:
2
max: E ðkÞ ðu, v, wÞ - bðkÞ
k=1
ð10:7Þ
s:t: E ðkÞ ðu, v, wÞ ≥ bðkÞ , k = 1, 2,
ðu, v, wÞ 2 F
where F is the set of (u, v, w) multipliers that can be used to measure the efficiencies
of the two divisions. Du et al. (2011) used the smallest possible efficiency of player
k to be b(k), in that b(1) and b(2) are measured from the worst conditions of (Xmax,
Zmin) and (Zmax, Ymin), respectively, where Xi max = max { Xij, j = 1, . . ., n},
Zg min = min { Zgj, j = 1, . . ., n}, Zg max = max { Zgj, j = 1, . . ., n}, and
Yr min = min { Yrj, j = 1, . . ., n}. And E ð1Þ ðu, v, wÞ = hg = 1 wg Z g0 = m i = 1 vi X i0
ð2Þ s h
and E ðu, v, wÞ = r = 1 ur Y r0 = g = 1 wg Z g0 are the efficiencies of the two divi-
sions. The system efficiency is the product of these two-division efficiencies. They
showed that if there is only one intermediate product, then their Nash bargaining
game yields the same results as applying the independent model to each division
separately.
Zhou et al. (2013) defined the breakdown point differently. First, Model (10.1) is
applied to obtain the system efficiency E0. Model (10.3) is then applied to obtain the
ð1Þþ ð2Þþ
largest possible efficiencies E0 and E 0 for divisions one and two, respectively.
ð1Þ - ð2Þþ ð2Þ -
The breakdown point is defined as bð1Þ = E 0 = E 0 /E 0 and b(2)= E 0 =E0/
ð1Þþ
E 0 . The DEA bargaining game under constant returns to scale is formulated as:
h s
g = 1 wg Z g0 ð1Þ - r = 1 ur Y r0 ð2Þ -
max: m - E0 × h
- E0
i = 1 vi X i0 g = 1 wg Z g0
h m
ð1Þ -
s:t: wg Z g0 ≥ E0 vi X i0
g=1 i=1
s h
ð2Þ -
ur Y r0 ≥ E 0 wg Z g0
r=1 g=1 ð10:8Þ
h m
wg Z gj - vi X ij ≤ 0, j = 1, . . . , n
g=1 i=1
s h
ur Y rj - wg Z gj ≤ 0, j = 1, . . . , n
r=1 g=1
ur , vi , wg ≥ 0, r = 1, . . . , s, i = 1, . . . , m, g = 1, . . . , h:
Zhou et al. (2013) showed that this linearly constrained nonlinear program pro-
vides a unique decomposition for the system efficiency. At optimality, we have
h m
ð1Þ ð1Þþ ð1 Þ -
E0 = wg Z g0 = vi X i0 = E0 × E0
g=1 i=1
s h
ð2Þ ð2Þþ ð2Þ -
E0 = ur Y r0 = wg Z g0 = E0 × E0
r=1 g=1
and
ð1Þ ð2Þ ð1Þþ ð1Þ - ð2Þþ ð2Þ - p

E0 × E0 = E0 × E0 × E0 × E0 = E0 × E0 = E0
For the basic two-stage systems, all of the existing studies on efficiency decom-
position show that the system efficiency is the product of the two-division efficien-
cies. Even for productivity changes over two periods, Kao and Hwang (2014) also
showed that the global Malmquist productivity index for the system is equal to the
product of those for the two divisions. This is an interesting property of the basic
two-stage system.
10.2.2 Efficiency Aggregation
The shortcoming of the independent model is that it is unable to measure the system
efficiency, although the efficiencies of the two divisions can be measured appropri-
ately. One way to address this is to define an aggregation of the division efficiencies
as the system efficiency and then use this as the objective function of a mathematical
program, using the operations of the two divisions as the constraints. After the
mathematical program is solved, the system and division efficiencies are obtained
at the same time. Different from efficiency decomposition where the system effi-
ciency is always expressed as the ratio of the aggregate exogenous output to the
aggregate exogenous input, efficiency aggregation has different ways to define the
system efficiency.
One straightforward way to aggregate the two-division efficiencies is to use their
average to be the system efficiency (Chiou et al. 2010). The input model under
variable returns to scale has the following form:
h
g = 1 wg Z g0 - w0
s
0 = max: 1 r = 1 ur Y r0 - u0
E m þ h
2 i = 1 vi X i0 g = 1 wg Z g0
h m
s:t: wg Z gj - w0 - vi X ij ≤ 0, j = 1, . . . , n
g=1 i=1
s h
ur Y rj - u0 - wg Z gj ≤ 0, j = 1, . . . , n
r=1 g=1
ur , vi , wg ≥ 0, r = 1, . . . , s, i = 1, . . . , m, g = 1, . . . , h:
ð10:9Þ

At optimality, h
g = 1 wg Z g0 - w0 = m
i = 1 vi X i0 and

s
r = 1 ur Y r0 - u0= h
g = 1 wg Z g0
are the efficiencies of divisions one and two,
respectively, and their average, which is the objective value E 0 , is the system
efficiency. This model can also be formulated from the output side, or the parameters
w0 and u0 can be deleted to yield a model under the assumption of constant returns to
scale. It is thus flexible for different management purposes. However, a weakness of
this model is the nonlinearity of the objective function, and thus a powerful computer
software package is needed to find a solution. This feature of nonlinearity also
prevents an analysis from the dual space.
A generalization of Model (10.9) is to use a weighted arithmetic average, with
pre-specified weights p(1) and p(2), of the two-division efficiencies as the objective
function. Chen et al. (2009a) defined the weight to be the proportion of the
aggregate input of the division in that of the two divisions, i.e.,
pð1Þ = m i = 1 vi X i0 =
m
i = 1 vi X i0 þ hg = 1 wg Z g0 Þ and pð2Þ = hg = 1 wg Z g0 =
m h
i = 1 vi X i0 þ g = 1 wg Z g0 : Since p(1) + p(2) = 1 and p(1), p(2) ≥ 0, the system
efficiency is a weighted arithmetic average of the two-division efficiencies. With
this definition of p(k), the system efficiency can be simplified as:
ð1Þ ð2Þ
pð1Þ × E0 þ pð2Þ × E0
h
g = 1 wg Z g0 - w0
m
i = 1 vi X i0
= m h
× m
i = 1 vi X i0 þ g = 1 wg Z g0 i = 1 vi X i0
h s
g = 1 wg Z g0 r = 1 ur Y r0 - u0
þ m h
× h
i = 1 vi X i0 þ g = 1 wg Z g0 g = 1 wg Z g0
h s
g = 1 wg Z g0 - w0 þ r = 1 ur Y r0 - u0
= m h
i = 1 vi X i0 þ g = 1 wg Z g0
The associated model is:
h s
ð g = 1 wg Z g0 - w0 Þ þ ð r = 1 ur Y r0 - u0 Þ
E 0 = max: m h
i = 1 vi X i0 þ g = 1 wg Z g0
h m
s:t: wg Z gj - w0 - vi X ij ≤ 0, j = 1, . . . , n
g=1 i=1
s h
ur Y rj - u0 - wg Z gj ≤ 0, j = 1, . . . , n
r=1 g=1
ur , vi , wg ≥ 0, r = 1, . . . , s, i = 1, . . . , m, g = 1, . . . , h:
ð10:10Þ
Note that since both the numerator and denominator have the common variables
wg, the objective function cannot be linearized by assigning one to the denominator
as a constraint, leaving the numerator as the objective function. It must be linearized
by applying the variable substitution technique introduced in Charnes and Cooper
(1962) for linear fractional programs, as described in the following.
Let the denominator of the objective function be equal to the inverse of a positive
m h
variable t, i.e., i = 1 vi X i0 þ g = 1 wg Z g0 = 1=t: In this case the objective func-
h s
tion becomes “max. g = 1 wg Z g0 - w0 þ r = 1 ur Y r0 - u0 t,” with the con-
m h
straint of i = 1 vi X i0 þ g = 1 wg Z g0 t = 1: Substituting ur = tur , u0 = tu0 ,
vi = tvi , wg = twg , and w0 = tw0 , where t > 0, Model (10.10), with the non-
Archimedean number ε as the lower bound for all multipliers restored, is then
linearized as:
h s
max: wg Z g0 - w0 þ ur Y r0 - u0
g=1 r=1
m h
s:t: vi X i0 þ wg Z g0 = 1
i=1 g=1
h m
wg Z gj - w0 - vi X ij ≤ 0, j = 1, . . . , n
g=1 i=1
s h
ur Y rj - u0 - wg Z gj ≤ 0, j = 1, . . . , n
r=1 g=1
t > 0, ur , vi , wg ≥ tε, r = 1, . . . , s, i = 1, . . . , m, g = 1, . . . , h:
Since smaller values for t obtain larger feasible regions to produce greater
efficiency scores, t can be ignored to leave the small non-Archimedean number ε
as the lower bound for the multipliers. After an optimal solution ðu , v , w , t Þ is
obtained, the objective value is the aggregate efficiency of the system, and the
ð1Þ h m
efficiencies of the two divisions are E 0 = g = 1 wg Z g0 - w0 = i = 1 vi X i0 and
ð2Þ s h
E0 = r = 1 ur Y r0 - u0 = g = 1 wg Z g0 :
The weights p(1) and p(2) are the most favorable ones for each DMU to measure its
system efficiency; and they may not be the same for other DMUs. Using the data in
Table 10.1 to illustrate this, the division and system efficiencies, and the associated
weights for cases of constant and variable returns to scale, are calculated via Model
(10.10), with the results shown in Table 10.3. It is noted first that the weights
selected by the DMUs are quite different, in both cases of constant and variable
returns to scale. Second, the system efficiency measured from the most favorable
weights is greater than that measured from Model (10.9), E 0 , with fixed weights of
0.5 for the two-division efficiencies for every DMU. Third, the division efficiencies
are exactly the same as those measured from the efficiency decomposition model
(10.1) under constant returns to scale. This is because there is only one intermediate
product, which, according to Liang et al. (2008), ensures the same result. However,
their system efficiencies are different, due to different definitions. Fourth, the
efficiencies of Division 1 are the same as those measured from the independent
model under variable returns to scale (referring to Table 10.1), while those of
Division 2 are not. This is because the efficiencies for Division 2 in Table 10.1 are
measured from the output side, and they are measured from the input side in Model
(10.10).
Model (10.10) is formulated from the input side. It can also be formulated from
the output side. Conventionally, the objective function of the output model is
ð1Þ
expressed as minimizing the inverse of the efficiency: 1=E0 = pð1Þ 1=E 0 þ
ð2Þ
pð2Þ 1=E 0 : This expression implies that the system efficiency is a weighted
harmonic average of the division efficiencies. Wang and Chin (2010) defined the
10.2
Table 10.3 Weighted average aggregation of efficiency measures for the example
Ratio-Form Efficiency Measures

ð1Þ (1) ð2Þ (2) ð1Þ ð2Þ
DMU E0 (p ) E0 (p ) E0 E0 E0 ( p(1)) E0 ( p(2)) E0 E0
A 1/2 (2/3) 2/5 (1/3) 7/15 0.45 1 (4/5) 1 (1/5) 1 1
B 1/2 (2/3) 2/5 (1/3) 7/15 0.45 5/8 (4/5) 2/3 (1/5) 19/30 0.6458
C 1 (1/2) 8/15 (1/2) 23/30 0.7667 1 (2/3) 2/3 (1/3) 8/9 0.8333
D 4/5 (5/9) 1 (4/9) 8/9 0.9 9/10 (5/11) 1 (6/11) 21/22 0.95
E 5/6 (6/11) 22/25 (5/11) 47/55 0.8567 1 (4/9) 1 (5/9) 1 1
235
weight associated with a division as the proportion of the aggregate output of this
division in that of the two divisions, and thus obtained an objective function of a
simpler form:
1 1 1
= pð1Þ × ð1Þ
þ pð2Þ × ð2Þ
E0 E0 E0
h m
g = 1 wg Z g0 i = 1 vi X i0 þ v0
= h s
× h
g = 1 wg Z g0 þ r = 1 ur Y r0 g = 1 wg Z g0
h
g = 1 wg Z g0 þ
s w0
r = 1 ur Y r0
þ h s
× s
g = 1 g g0 þ
w Z r = 1 ur Y r0 r = 1 ur Y r0
m h
i = 1 vi X i0 þ v0 þ g = 1 wg Z g0 þ w0
= h s
g = 1 wg Z g0 þ r = 1 ur Y r0
This fractional objective function can be linearized similarly as described after

Model (10.10).
10.3 Distance Function Efficiency Measures
There are three types of distance parameters that have appeared in the literature:
system, division, and factor. The first type has only one parameter attached to either
the inputs or the outputs and is referred to as the input- or output-oriented model. The
second type has two parameters, one attached to the first and the other attached to the
second division. The third type has m + s parameters attached to the m input and
s output factors. Since the model corresponding to the third type is equivalent to the
slacks-based efficiency measures, it will be discussed in Sect. 10.4. This section
discusses the other two types.
10.3.1 System Parameter
The distance function defined by Shephard (1970) for measuring the efficiency of the
system as a whole unit was extended by Färe and Grosskopf (2000) to measure the
efficiency of network systems. For a two-stage system with the structure shown in
ð1Þ
Fig. 10.1, the production possibility set is: T = X, Y, Z, Z j nj = 1 λj X ij ≤ X i ,
n ð1Þ n ð2Þ
i = 1,. . ., m, j = 1 λj Z gj ≥ Z g , g = 1,. . ., h, j = 1 λj Z gj ≤ Z g , g = 1,. . ., h,
n ð2Þ ðk Þ
j = 1 λj Y rj ≥ Y r , r = 1,. . ., s, λj ≥ 0, j = 1,. . ., n, k = 1, 2}. Under variable
n ðk Þ
returns to scale, the constraint j = 1 λj = 1 is added. The input distance parameter
10.3 Distance Function Efficiency Measures 237
is the minimum contraction factor that can keep the inputs of the DMU being
evaluated in the production possibility set. The mathematical program for finding
the value of this parameter is:
min: θ
n
ð1Þ
s:t: λj X ij ≤ θX i0 , i = 1, . . . , m
j=1
n
ð1Þ
λj Z gj ≥ Z g0 , g = 1, . . . , h
j=1
n
ð2Þ
ð10:11Þ
λj Z gj ≤ Z g0 , g = 1, . . . , h
j=1
n
ð2Þ
λj Y rj ≥ Y r0 , r = 1, . . . , s
j=1
ðkÞ
λj ≥ 0, j = 1, . . . , n, k = 1, 2
Similar to the ratio-form models, the non-Archimedean number ε has been

ignored for simplicity of expression.
Compared to the black-box model of
min: θ
n
s:t: λj X ij ≤ θX i0 , i = 1, . . . , m
j=1
n ð10:12Þ
λj Y rj ≥ Y r0 , r = 1, . . . , s
j=1
λj ≥ 0, j = 1, . . . , n
ðk Þ
the network model (10.11) has one more set of variables, λj , which, on the one
hand, makes the model less stringent and results in a smaller efficiency measure. It
also has two more sets of constraints, associated with the intermediate products Zg0,
which, on the other hand, make the model more stringent, resulting in a larger
efficiency. When these two effects are combined, it is not guaranteed that the
network model will obtain a smaller efficiency than that measured from the black-
box model.
The efficiency decomposition model (10.1), which is in ratio form, is able to
measure the system and division efficiencies. This model’s dual, when the redundant
system constraints are ignored, is:
min: θ
n
ð1Þ
s:t: λj X ij ≤ θX i0 , i = 1, . . . , m
j=1
n n
ð1Þ ð2Þ
λj Z gj ≥ λj Z gj , g = 1, . . . , h ð10:13Þ
j=1 j=1
n
ð2Þ
λj Y rj ≥ Y r0 , r = 1, . . . , s
j=1
ðkÞ
λj ≥ 0, j = 1, . . . , n, k = 1, 2
This model is less stringent than the Färe-Grosskopf distance function model
(10.11), because its second set of constraints is the aggregation of the second and
third sets of constraints of Model (10.11) based on the transitivity relationship via
Zg0. The system efficiency measured from this model is thus less than or equal to that
measured from Model (10.11). The projection points for Xi0, Zg0, and Yr0, based on
the constraints of Model (10.13), are:
^ i0 =
n
ð1Þ
X λ j X ij , i = 1, . . . , m
j=1
ðoutÞ ð1Þ
Z^ g0 =
n
λ j Z gj , g = 1, . . . , h
ðinÞ
j=1
ð2Þ
ð10:14Þ
Z^ g0 =
n
λ j Z gj , g = 1, . . . , h
j=1
ð2Þ
Y^ r0 =
n
λ j Y rj , r = 1, . . . , s
j=1
ðoutÞ ðinÞ
where Z g0 and Z g0 are the projection points for Zg0 as the output of Division 1 and
n n ð1Þ ð2Þ
input of Division 2, respectively. The constraint j = 1 λj Z gj ≥ j = 1 λj Z gj in
Model (10.13) ensures that the intermediate product as an output is greater than or
equal to that as an input. This indicates that an assumption of strong disposability is
imposed. If the amount produced by Division 1 is larger than that needed by Division
2, then the excessive amount is discarded.
Chen et al. (2013) stressed that the ratio model is able to measure the division
efficiencies while the envelopment model, in which the distance function efficiency
measure is expressed, is able to show the projections, or targets, of the factors. This is
especially important for network systems, because an intermediate product produced
by a division is expected to have larger values to be more efficient for this division
and is expected to have smaller values to be more efficient for the division that uses it
for production. The supply and demand divisions of the same intermediate product
thus have conflicting goals regarding the value of this intermediate product. In the
study of Lim and Zhu (2016), the primal-dual relationship of models (10.1) and
Table 10.4 Projections of the five DMUs measured from different models
Input model Output model
ðoutÞ ðinÞ ðoutÞ ðinÞ
DMU θ X Z Z Y φ X Z Z Y
A 1/5 0.4 0.4 0.4 0.5 5 2 2 2 2.5
B 1/5 0.8 0.8 0.8 1 5 4 4 4 5
C 8/15 1.6 1.6 1.6 2 15/8 3 3 3 3.75
D 4/5 4 4 4 5 5/4 5 5 5 6.25
E 11/15 4.4 4.4 4.4 5.5 15/11 6 6 6 7.5
(10.13) was applied to measure the division efficiencies and derive the projections in
one model.
Model (10.13) is formulated from the input side. The following model is formu-
lated from the output side:
max: φ
n
ð1Þ
s:t: λj X ij ≤ X i0 , i = 1, . . . , m
j=1
n n
ð1Þ ð2Þ
λj Z gj ≥ λj Z gj , g = 1, . . . , h ð10:15Þ
j=1 j=1
n
ð2Þ
λj Y rj ≥ φY r0 , r = 1, . . . , s
j=1
ðkÞ
λj ≥ 0, j = 1, . . . , n, k = 1, 2
The system efficiency is the inverse of the objective function, 1/φ, and the targets
have the same equations as those of Expression (10.14). Table 10.4 shows the
objective values and projections of the five DMUs solved from Models (10.13)
ðoutÞ ðinÞ
and (10.15). In this example Z is the same as Z for all DMUs for each of the
two models. However, they are different between the two models. For example, the
projection of DMU D is Z D = 4 for the input model and is Z D = 5 for the output
model. It is also noted that the efficiency measured from the output model is the
inverse of the input model. This property always holds for the basic two-stage
system and can be proved easily from their corresponding ratio models.
ð1Þ
Some studies, for example, Chen et al. (2010), use the constraints nj = 1 λj Z gj ≥
ð2Þ
Z g and nj = 1 λj Z gj ≤ Z g to find the same target zg for the two divisions, so there will
not be conflict between them.
10.3.2 Division Parameters
The distance function can be defined either from the input side to minimize the input
contraction factor or from the output side to maximize the output expansion factor.
One approach is to combine minimizing the input contraction factor of the first
division with maximizing the output expansion factor of the second division. The
model proposed by Chen and Zhu (2004) under constant returns to scale is:
min: θ - φ
n
ð1Þ
s:t: λj X ij ≤ θX i0 , i = 1, . . . , m
j=1
n n
ð1Þ ð2Þ
λj Z gj ≥ λj Z gj , g = 1, . . . , h
j=1 j=1 ð10:16Þ
n
ð2Þ
λj Y rj ≥ φY r0 , r = 1, . . . , s
j=1
θ ≤ 1, φ ≥ 1
ðkÞ
λj ≥ 0, j = 1, . . . , n, k = 1, 2
It is expected that, at optimality, θ* and φ* will be the efficiencies of the first and
second divisions, respectively. To verify this, we replace λj = θμj , λj = θμj ,
and φ = θη in this model to obtain the following equivalent model:
max: θðη - 1Þ
n
ð1Þ
s:t: μj X ij ≤ X i0 , i = 1, . . . , m
j=1
n n
ð1Þ ð2Þ
μj Z gj ≥ μj Z gj , g = 1, . . . , h
j=1 j=1 ð10:17Þ
n
ð2Þ
μj Y rj ≥ ηY r0 , r = 1, . . . , s
j=1
θ ≤ 1, θη ≥ 1
ðkÞ
μj ≥ 0, j = 1, . . . , n, k = 1, 2
Note that the original objective function of “min. θ -φ ” has been changed to
“max. φ- θ” for variable substitutions. At optimality, θ must be equal to 1. For if θ
is strictly less than 1, then the constraint θη ≥ 1 implies that η will be strictly greater
than 1. In this case we can define θ = θ þ δ, where δ is a very small positive
number, such that θ = θ þ δ ≤ 1 and θη = ðθ þ δÞη > θ η ≥ 1: In other words,
θ, η , μ is a feasible solution to Model (10.17). The objective value of this
solution is θðη - 1Þ = ðθ þ δÞðη - 1Þ > θ ðη - 1Þ: Since a better objective
Table 10.5 Data and effi- ð1Þ ð2Þ

DMU X Z Y θ* φ* E0 E0 E0
ciency measures for the
division-parameter example A 1 2 1 1 2 1 1/2 1/2
B 2 3 3 1 4/3 3/4 1 3/4
C 3 4 4 1 3/2 2/3 1 2/3
D 4 6 5 1 8/5 3/4 5/6 5/8
value is obtained, contradicting that θ < 1 is optimal, θ must thus be equal to 1 to

be an optimal solution.
By requiring θ = 1, Model (10.17) is simplified to:
max: η - 1
n
ð1Þ
s:t: μj X ij ≤ X i0 , i = 1, . . . , m
j=1
n n
ð1Þ ð2Þ
μj Z gj ≥ μj Z gj , g = 1, . . . , h
j=1 j=1 ð10:18Þ
n
ð2Þ
μj Y rj ≥ ηY r0 , r = 1, . . . , s
j=1
η≥1
ðkÞ
μj ≥ 0, j = 1, . . . , n, k = 1, 2
The constant -1 in the objective function has no effect on determining the

optimal solution and can be deleted. The resulting model is exactly the same as
Model (10.15), the dual of the efficiency decomposition model of Kao and Hwang
(2008) from the output side, in that η is the system efficiency.
This discussion indicates that the optimal θ and φ in Model (10.16) do not
represent the efficiencies of the two divisions. In fact, θ is always equal to 1 and φ
represents the efficiency of the system. This is a pitfall in modeling the basic
two-stage systems (Chen et al., 2009b).
As stressed in Chen et al. (2013), the efficiency of each stage must be measured
via the multiplier-form model. To see this, consider an example of four DMUs using
one-input X in Division 1 to produce one intermediate product Z, which in turn is
used in Division 2 to produce one output Y, with the data shown in Table 10.5, and
the DMUs depicted in Fig. 10.4. By applying Model (10.16), the optimal values for θ
and φ are solved, as shown in columns five and six of Table 10.5.
As expected, all DMUs have a value of one for θ*. To verify that φ* is the
efficiency of the system (in reciprocal form), we formulate the dual of Model (10.16)
as follows:
max: e2 - e1
h m
s:t: wg Z gj - vi X ij ≤ 0, j = 1, . . . , n
g=1 i=1
s h
ur Y rj - wg Z gj ≤ 0, j = 1, . . . , n
r=1 g=1
ð10:19Þ
m
vi X i0 - e1 = 1
i=1
s
ur Y r0 - e2 = 1
r=1
ur , vi , wg ≥ 0, r = 1, . . . , s, i = 1, . . . , m, g = 1, . . . , h:
ð1Þ
At optimality, the division efficiencies, based on the constraints, areE0 =
h m ð2Þ s h
g = 1 wg Z g0 = i = 1 vi X i0 and E 0 = r = 1 ur Y r0 = g = 1 wg Z g0 , and the system effi-
ð1Þ ð2Þ
ciency is E 0 = sr = 1 ur Y r0 = m
i = 1 vi X i0 = E 0 × E 0 : By applying this model to the
ð1Þ ð2Þ
data contained in Table 10.5, the values of E0 , E 0 , and E0 are calculated, as shown
ð1Þ ð2Þ
in the last three columns. The values of E 0 and E 0 can be checked graphically
from Fig. 10.4. It is noted that the value of φ , in column six, is exactly the inverse of
*
E0, in the last column, for every DMU.
D ( 2) 6 D (1)
4 C (1)
C ( 2)
B (1)
B ( 2)
A( 2) 2 A(1)
Y X
6 4 2 O 2 4
Fig. 10.4 Frontiers of the two stages of the division-parameter example

10.4 Slacks-Based Efficiency Measures 243
10.4 Slacks-Based Efficiency Measures
Suppose the input and output distance parameters are allowed to be different for each
factor. A model for measuring the system efficiency under constant returns to scale
can be formulated as (Yu 2012):
1 m
θ
i=1 i
min: m
1 s
φ
r=1 r
s
n
ð1Þ
s:t: λj X ij ≤ θi X i0 , i = 1, . . . , m
j=1
n n
ð1Þ ð2Þ
λj Z gj ≥ λj Z gj , g = 1, . . . , h ð10:20Þ
j=1 j=1
n
ð2Þ
λj Y rj ≥ φr Y r0 , r = 1, . . . , s
j=1
θi ≤ 1, i = 1, . . . , m
φr ≥ 1, r = 1, . . . , s
ðkÞ
λj ≥ 0, j = 1, . . . , n, k = 1, 2
ðk Þ
The constraint nj = 1 λj = 1 is added if the kth division is assumed to have a
variable returns to scale technology. The inequality constraints corresponding to the
ð1Þ ð2Þ
intermediate products, nj = 1 λj Z gj ≥ nj = 1 λj Z gj , can be set to equality if “fixed
link” or “cooperation” between the two divisions is assumed. This model is basically
the enhanced Russell efficiency measure of Pastor et al. (1999) applied to the basic
two-stage system. By defining θi = X i0 - si- =X i0 as the input efficiency of the ith
input and ϕr = Y r0 þ sþ r =Y r0 as the (inverse) output efficiency of the rth output,
Model (10.20) becomes:
1 m
1- i=1
ðsi- =X i0 Þ
min: m
1 s
1þ r=1
ðsþ
r =Y r0 Þ
s
n
ð1Þ
s:t: λj X ij þ si- = X i0 , i = 1, . . . , m
j=1
n n ð10:21Þ
ð1Þ ð2Þ
λj Z gj ≥ λj Z gj , g = 1, . . . , h
j=1 j=1
n
ð2Þ
λj Y rj - sþ
r = Y r0 , r = 1, . . . , s
j=1
ðkÞ
λj ≥ 0, j = 1, . . . , n, k = 1, 2
The objective function is the product of the average input and average output
efficiencies. This model is essentially the network slacks-based model of Tone and
Tsutsui (2009) for basic two-stage systems.
Model (10.21) can be classified as an efficiency decomposition one, because the
objective function only involves the slacks related to the exogenous inputs and final
outputs. If it also contains those related to the intermediate products, then it will be
an efficiency aggregation model.
A similar idea to Model (10.20) is to use a direction vector g to replace (X0, Y0) in
the objective function. As discussed in Chap. 6, direction vectors are useful in
measuring efficiencies with undesirable factors. Let Blj be the lth undesirable output,
l = 1, . . ., t, of the jth DMU. The model proposed by Fukuyama and Weber (2010) to
handle undesirable outputs is:
1 1
m
si- 1 sþ
s
1
t
sol
min: þ r
Y þ t
3 m gX s
i=1 i r=1
g r gB
l=1 l
n
ð1Þ
s:t: λj X ij þ si- = X i0 , i = 1, . . . , m
j=1
n n
ð1Þ ð2Þ
λj Z gj ≥ λj Z gj , g = 1, . . . , h
j=1 j=1 ð10:22Þ
n
ð2Þ
λj Y rj - sþ
r = Y r0 , r = 1, . . . , s
j=1
n
ð2Þ
λj Blj þ sol = Bl0 , l = 1, . . . , q
j=1
ðkÞ
λj ≥ 0, j = 1, . . . , n, k = 1, 2
Each term of (slack)/(direction) in the objective function shows the relative

distance between the DMU and its target of the corresponding factor. The objective
function merely provides a mechanism for finding a target for making improve-
ments, whose value does not represent the relative level of inefficiency. A value of
zero indicates that the DMU is efficient, while a positive value indicates inefficiency.
Since the largest value of the objective function is not known, a relative scale
showing the extent of inefficiency is unavailable, although it is known that DMUs
with larger objective values are less efficient. Similar to other envelopment models,
the constraints of Model (10.22) are able to show the targets for a DMU to become
efficient. This model was used by Akther et al. (2013) to measure the performance of
21 commercial banks in Bangladesh.
The slacks-based measures have received relatively little attention in the basic
two-stage system. We will discuss this type of model in more detail in the following
chapters, where it appears more in the related literature.
The independent model was the only choice for studying basic two-stage systems
before the network DEA models were proposed. Moreover, even with the develop-
ment of network models, the independent model has remained in use. Two of the
major applications were investigating the effects of IT on the performance of
22 banks with the two stages of fund collection and investment (Wang et al. 1997)
and examining the performance of the top 55 US commercial banks with respect to
profitability and marketability (Seiford and Zhu 1999). Based on these works, other
studies have been carried out examining Fortune 500 companies (Zhu 2000),
245 banks in the USA with assets greater than one billion US dollars (Luo 2003),
14 financial holding companies in Taiwan (Lo and Lu 2006, 2009), US S&P
500 firms (Lo 2010), 40 IC fabless firms in Taiwan (Lu and Hung 2009), 13 com-
mercial banks in Greece (Tsolas 2011), and 19 construction firms in Greece (Tsolas
2013). Other similar studies include the productivity of 13 retail stores in the USA
(Keh and Chu 2003), the front-office and on-field operations of 30 teams in two
Major League Baseball leagues (Sexton and Lewis 2003), the performance of
manager- and physician-controlled processes in physician care (Chilingerian and
Sherman 2004), the flexibility and execution competence of 70 cases from the
National Association of Purchasing Managers (Narasimhan et al. 2004), and the
production and inventory efficiencies of 72 life and health insurance companies in
Canada (Yang 2006), the marketing and services performance of 49 hotels in an
Asia-Pacific hotel chain (Keh et al. 2006), the operating efficiency and effectiveness
of 28 online stockbrokers in Taiwan (Ho and Oh 2008), the R&D invention
efficiency and marketability of 28 Taiwanese IC-design firms (Lu 2009), and the
operational and profitability efficiencies of 11 life insurance companies in Taiwan
(Tsai and Wang 2010).
Many models have been proposed with regard to the ratio-form efficiency
measure after the work presented in Kao and Hwang (2008), along with various
applications of this approach, such as those examining the marketability and prof-
itability of 40 internet companies (Cao and Yang 2011), the production and profit-
ability performance of 23 printed circuit board manufacturing firms in Taiwan (Liu
and Wang 2009), the efficiency of 55 branches of two banks in Taiwan (Yang and
Liu 2012), and the design performance of 534 carlines, in which the first stage is an
industrial design module and the second a bio design one (Chen et al. 2012). While
Kao and Hwang (2008) assumed complete allocative efficiency, Sahoo et al. (2014)
used a single network technology to decompose the system efficiency. Kao and Liu
(2011) discussed how to calculate efficiency when the observations were fuzzy data.
For efficiency aggregation, Lim and Zhu (2013) made some modifications to the
ideas presented in Chiou et al. (2010) which used the average of the two-division
efficiencies as the criterion to measure the technical and scale efficiencies of
39 intercity bus companies in Taiwan. Lu et al. (2012) proposed using a BCC
model for each division to measure the production and marketing efficiencies of
30 airline companies in the USA. Wang et al. (2014) measured the efficiency of
16 major Chinese commercial banks with undesirable outputs. There are many
studies investigating the decomposition of the system efficiencies in the basic two-
stage system (Despotis et al., 2016; Guo et al., 2017; Li, 2017).
The distance function efficiency measure has also been widely studied for the
basic two-stage system. Yang et al. (2011) measured the supply chain efficiency of
17 bank branches of the China Construction Bank in Anhui Province, China. This
model has also been extended to include two more objectives, in addition to
maximizing the output distance parameter, in a goal programming setting by Sheth
et al. (2007) in studying 60 bus routes in the Christiansburg-Blacksburg region in
Virginia, USA, from the perspectives of both provider and passenger. Lewis et al.
(2013) proposed an iterative method of minimizing the input parameter θ and
maximizing the output parameter φ at the same time. Rho and An (2007) modified
the model of Chen and Zhu (2004) to identify weakly efficient DMUs. A variation of
this model was used by Holod and Lewis (2011) to study the role of deposits in the
bank production process, as whether these should be treated as an input or output.
Saranga and Moser (2010) evaluated the purchasing and supply management per-
formance of 120 international firms with turnovers of more than three billion US
dollars. Chiu and Huang (2011) measured the performance of 57 international tourist
hotels in Taiwan. Liu and Lu (2012) used centrality measures for ranking in a study
similar to that of Liu et al. (2010), in which the R&D efficiencies of two processes at
32 research institutes in Taiwan, technology development and technology diffusion,
were measured independently. Azadi et al. (2014) developed two DEA models to set
targets for 24 bus companies in Tehran to meet green supply chain management’s
requirement.
For the slacks-based approach, Zhu et al. (2014) measured the eco-efficiency of
10 pesticides in the two stages of production and usage of the pesticide. Lu et al.
(2014) investigated the performance of the national innovation system of 30 coun-
tries from the aspects of R&D and economic efficiencies. Liu et al. (2015) proposed
a model to deal with undesirable inputs, outputs, and intermediate products. A result
similar to that described in Chen et al. (2013) that a multiplier model is required to
find the division efficiencies and an envelopment model is needed to determine the
targets for inefficient DMU was also obtained in Chen et al. (2016) based on a
slacks-based measure.
References
Akther, S., Fukuyama, H., & Weber, W. L. (2013). Estimating two-stage network slacks-based
inefficiency: An application to Bangladesh banking. Omega, 41, 88–96.
Azadi, M., Shabani, A., Khodakarami, M., & Saen, R. F. (2014). Planning in feasible region by
two-stage target-setting DEA methods: An application in green supply chain management of
public transportation service providers. Transportation Research Part E- Logistics and Trans-
portation Review, 70, 324–338.
Cao, X., & Yang, F. (2011). Measuring the performance of internet companies using a two-stage
data envelopment analysis model. Enterprise Information Systems, 5, 207–217.
References 247
Report CCS #532, Center for Cybernetic Studies, University of Texas-Austin, Austin, TX.
Chen, C., Zhu, J., Yu, J. Y., & Noori, H. (2012). A new methodology for evaluating sustainable
product design performance with two-stage network data envelopment analysis. European
Chen, Y., Cook, W. D., Kao, C., & Zhu, J. (2013). Network DEA pitfalls: Divisional efficiency and
frontier projection under general network structures. European Journal of Operational
Research, 226, 507–515.
Chen, Y., Cook, W. D., Li, N., & Zhu, J. (2009a). Additive efficiency decomposition in two-stage
DEA. European Journal of Operational Research, 196, 1170–1176.
Chen, Y., Liang, L., & Zhu, J. (2009b). Equivalence in two-stage DEA approaches. European
Chen, Y., Cook, W. D., & Zhu, J. (2010). Deriving the DEA frontier for two-stage processes.
Chen, Y., Li, Y., Liang, L., Salo, A., & Wu, H. (2016). Frontier projection and efficiency
decomposition in two-stage processes with slacks-based measures. European Journal of Oper-
ational Research, 250(2), 543–554.
Chen, Y., & Zhu, J. (2004). Measuring information technology’s indirect impact on firm perfor-
mance. Information Technology and Management, 5, 9–22.
Chilingerian, J. A., & Sherman, D. H. (2004). Health care applications: From hospitals to physi-
cians, from productive efficiency to quality frontiers. In W. W. Cooper, L. M. Seiford, & J. Zhu
(Eds.), Handbook of data envelopment analysis (pp. 481–537). Springer.
Chiou, Y. C., Lan, L. W., & Yen, B. T. H. (2010). A joint measurement of efficiency and
effectiveness for non-storable commodities: Integrated data envelopment analysis approaches.
Chiu, Y. H., & Huang, C. W. (2011). Evaluating the optimal occupancy rate, operational efficiency,
and profitability efficiency of Taiwan’s international tourist hotels. Service Industries Journal,
31, 2145–2163.
Cron, W., & Sobol, M. (1983). The relationship between computerization and performance: A
strategy for maximizing economic benefits of computerization. Information and Management,
6, 171–181.
Despotis, D. K., Koronakos, G., & Sotiros, D. (2016). The “weak-link” approach to network DEA
for two-stage processes. European Journal of Operational Research, 254, 481–492.
Du, J., Liang, L., Chen, Y., Cook, W. D., & Zhu, J. (2011). A bargaining game model for measuring
performance of two-stage network structures. European Journal of Operational Research, 210,
390–397.
Fukuyama, H., & Weber, W. L. (2010). A slacks-based inefficiency measure for a two-stage system
with bad outputs. Omega, 38, 398–409.
Guo, C., Shureshjani, R. A., Foroughi, A. A., & Zhu, J. (2017). Decomposition weights and overall
efficiency in two-stage additive network DEA. European Journal of Operational Research,
257, 896–906.
Ho, C. T. B., & Oh, K. B. (2008). Measuring online stockbroking performance. Industrial
Management & Data Systems, 108, 988–1004.
Holod, D., & Lewis, H. F. (2011). Resolving the deposit dilemma: A new DEA bank efficiency
model. Journal of Banking & Finance, 35, 2801–2810.
Research, 185, 418–429.
Kao, C., & Hwang, S. N. (2011). Decomposition of technical and scale efficiencies in two-stage
production systems. European Journal of Operational Research, 211, 515–519.
Kao, C., & Hwang, S. N. (2014). Multi-period efficiency and Malmquist productivity index in
two-stage production systems. European Journal of Operational Research, 232, 512–521.
Kao, C., & Liu, S. T. (2011). Efficiencies of two-stage systems with fuzzy data. Fuzzy Sets and
Systems, 176, 20–35.
Keh, H. T., & Chu, S. F. (2003). Retail productivity and scale economies at the firm level: A DEA
Keh, H. T., Chu, S. F., & Xu, J. Y. (2006). Efficiency, effectiveness and productivity of marketing
in services. European Journal of Operational Research, 170, 265–276.
Lewis, H. F., Mallikarjun, S., & Sexton, T. R. (2013). Unoriented two-stage DEA: The case of the
oscillating intermediate products. European Journal of Operational Research, 229, 529–539.
Li, X. (2017). A fair evaluation of certain stage in a two-stage structure: Revisiting the typical
two-stage DEA approaches. Omega, 68, 155–167.
Liang, L., Cook, W. D., & Zhu, J. (2008). DEA models for two-stage processes: Game approach
and efficiency decomposition. Naval Research Logistics, 55, 643–653.
Lim, S., & Zhu, J. (2013). Integrated data envelopment analysis: Global vs. local optimum.
Lim, S., & Zhu, J. (2016). A note on two-stage network DEA model: Frontier projection and
duality. European Journal of Operational Research, 248, 342–346.
Liu, J. S., & Lu, W. M. (2012). Network-based method for ranking of efficient units in two-stage
DEA models. Journal of the Operational Research Society, 63, 1153–1164.
Liu, J. S., Lu, W. M., & Lin, B. J. Y. (2010). DEA and ranking with the network-based approach: A
case of R&D performance. Omega, 38, 453–464.
Liu, S. T. (2011). A note on efficiency decomposition in two-stage data envelopment analysis.
Liu, S. T., & Wang, R. T. (2009). Efficiency measures of PCB manufacturing firms using relational
two-stage data envelopment analysis. Expert Systems with Applications, 36, 4935–4939.
Liu, W. B., Zhou, Z. B., Ma, C. Q., Liu, D. B., & Shen, W. F. (2015). Two-stage DEA models with
undesirable input-intermediate-outputs. Omega, 56, 74–87.
Lo, S. F. (2010). Performance evaluation for sustainable business: A profitability and marketability
framework. Corporate Social Responsibility and Environmental Management, 17, 311–319.
Lo, S. F., & Lu, W. M. (2006). Does size matter? Finding the profitability and marketability
benchmark of financial holding companies. Asia-Pacific Journal of Operational Research, 23,
229–246.
Lo, S. F., & Lu, W. M. (2009). An integrated performance evaluation of financial holding
companies in Taiwan. European Journal of Operational Research, 198, 341–350.
Lu, W. C. (2009). The evolution of R&D efficiency and marketability: Evidence from Taiwan's
IC-design industry. Asian Journal of Technology Innovation, 17, 1–26.
Lu, W. M., & Hung, S. W. (2009). Evaluating profitability and marketability of Taiwan's IC fabless
firms: A DEA approach. Journal of Scientific and Industrial Research, 68, 851–857.
Lu, W. M., Kweh, Q. L., & Huang, C. L. (2014). Intellectual capital and national innovation
systems performance. Knowledge-Based Systems, 71, 201–210.
Lu, W. M., Wang, W. K., Hung, S. W., & Lu, E. T. (2012). The effects of corporate governance on
airline performance: Production and marketing efficiency perspectives. Transportation
Research Part E-Logistics and Transportation Review, 48, 529–544.
Luo, X. M. (2003). Evaluating the profitability and marketability efficiency of large banks: An
application of data envelopment analysis. Journal of Business Research, 56, 627–635.
Narasimhan, R., Talluri, S., & Das, A. (2004). Exploring flexibility and execution competencies of
manufacturing firms. Journal of Operations Management, 22, 91–106.
Nash, J. F. (1950). The bargaining problem. Econometrica, 18, 155–162.
Nash, J. F. (1953). Two-person cooperative games. Econometrica, 21, 128–140.
References 249
Pastor, J. T., Ruiz, J. L., & Sirvent, I. (1999). An enhanced DEA Russell graph efficiency measure.
Rho, S., & An, J. (2007). Evaluating the efficiency of a two-stage production process using data
envelopment analysis. International Transactions in Operational Research, 14, 395–410.
Sahoo, B. K., Zhu, J., Tone, K., & Klemen, B. M. (2014). Decomposing technical efficiency and
scale elasticity in two-stage network DEA. European Journal of Operational Research, 233,
584–594.
Saranga, H., & Moser, R. (2010). Performance evaluation of purchasing and supply management
using value chain DEA approach. European Journal of Operational Research, 207, 197–205.
Seiford, L. M., & Zhu, J. (1999). Profitability and marketability of the top 55 US commercial banks.
Sexton, T. R., & Lewis, H. F. (2003). Two-stage DEA: An application to Major League Baseball.
Sheth, C., Triantis, K., & Teodorović, D. (2007). Performance evaluation of bus routes: A provider
and passenger perspective. Transportation Research Part E-Logistics and Transportation
Review, 43, 453–478.
Tsai, H. Y., & Wang, A. S. (2010). The efficiency analysis of life insurance company in Taiwan:
Two-stage data envelopment analysis. Journal of Testing and Evaluation, 38, 283–290.
Tsolas, I. E. (2011). Relative profitability and stock market performance of listed commercial banks
on the Athens exchange: A non-parametric approach. IMA Journal of Management Mathemat-
ics, 22, 323–342.
Tsolas, I. E. (2013). Modeling profitability and stock market performance of listed construction
firms on the Athens exchange: Two-stage DEA approach. Journal of Construction Engineering
and Management, 139, 111–119.
Wang, C. H., Gopal, S., & Zionts. (1997). Use of data envelopment analysis in assessing informa-
tion technology impact on firm performance. Annals of Operations Research, 73, 191–213.
Wang, K., Huang, W., & Wu, J. (2014). Efficiency measures of the Chinese commercial banking
system using an additive two-stage DEA. Omega, 44, 5–20.
Wang, Y. M., & Chin, K. S. (2010). Some alternative DEA models for two-stage process. Expert
Systems with Applications, 37, 8799–8808.
Yang, C., & Liu, H. M. (2012). Managerial efficiency in Taiwan bank branches: A network DEA.
Economic Modelling, 29, 450–461.
Yang, F., Wu, D., Liang, L., Bi, G., & Wu, D. D. (2011). Supply chain DEA: Production possibility
set and performance evaluation model. Annals of Operations Research, 185, 195–211.
Yang, Z. (2006). A two-stage DEA model to evaluate the overall performance of Canadian life and
health insurance companies. Mathematical and Computer Modelling, 43, 910–919.
Yu, M. M. (2012). Performance assessment of transport services with the ERM-NDEA model:
Evidence from a domestic airline in Taiwan. Transportation Planning and Technology, 35,
697–714.
Zhou, Z., Sun, L., & Yang, W. (2013). A bargaining game model for efficiency decomposition in
the centralized model of two-stage systems. Computers & Industrial Engineering, 64, 103–108.
Zhu, J. (2000). Multi-factor performance measure model with an application to fortune 500 com-
panies. European Journal of Operational Research, 123, 105–124.
Zhu, Z., Wang, K., & Zhang, B. (2014). Applying a network data envelopment analysis model to
quantify the eco-efficiency of products: A case study of pesticides. Journal of Cleaner Produc-
tion, 69, 67–73.
Chapter 11
General Two-Stage Systems
The basic two-stage system discussed in the preceding chapter describes a produc-
tion type where all intermediate products produced by the first division are con-
sumed by the second division for production. Specifically, no intermediate products
flow out of the system, and the second division does not consume other inputs
supplied from outside, except for the intermediate products. However, it should be
noted that real-world cases are usually more complicated than this basic two-stage
system. For example, some intermediate products may flow out of the system to be
sold as spare parts, and the second division may need workers to work on the
intermediate products to become the final products. Taking these situations into
account, we then have a general two-stage system, which allows the first division to
have final outputs and the second division to have exogenous inputs. Several models
have been proposed for measuring the efficiency of this type of system, and many
applications have been reported in the literature (Kao 2014a).
Theoretically, the complete general two-stage system is more complicated, in that
not only can some outputs produced by the first division be supplied to the second
division for production, but some outputs produced by the second division can also
be used by the first. The latter case is called feedback. To make the discussion
clearer, we separate the complete general two-stage structure into two, the basic two-
stage system with feedback and the general two-stage system without feedback,
although they can be discussed together. The discussion of this chapter starts with
the feedback system. The models that have been proposed to measure the efficiency
of the general two-stage system are then introduced, including independent, ratio-
form, distance function, and slacks-based efficiency measures. After these, a special
topic, shared input, which concerns the allocation of some common inputs to the two
divisions that will bring the largest benefits to the system, is discussed. Finally, some
applications are briefly described to give some idea of how the general two-stage
system can be applied to real-world problems.
https://doi.org/10.1007/978-3-031-27593-7_11
252 11 General Two-Stage Systems
11.1 Feedback System
The feedback system discussed in this section is a basic two-stage one, except that
some outputs of the second division are fed back to the first as a part of the inputs for
production. The first study of feedback systems in DEA is Liang et al. (2011), which
examines the performance of a university, in which the final output, research funds,
is fed back from the second stage to the first, together with other exogenous inputs, to
produce journal articles, which in turn generate research funds in the second stage.
Other prominent examples of the feedback system are the recycling of waste
materials and wastewater.
The feedback system has the structure depicted in Fig. 11.1, where the inputs Xi,
i = 1, . . ., m, are supplied to the first division from outside, together with some
factors Fl, l = 1, . . ., q, fed back from the second division, to produce the interme-
diate products Zg, g = 1, . . ., h, for the second division to produce the final outputs
Yr, r = 1, . . ., s, and the feedback outputs Fl. The inputs consumed by Division 1 are
the exogenous inputs Xi and the feedback outputs Fl, and the outputs produced are
the intermediate products Zg. The efficiency of this division can thus be measured as
ð1Þ q
E 0 = hg = 1 wg Z g0 = m
i = 1 vi X i0 þ l = 1 cl F l0 : For Division 2 the intermediate
products Zg are consumed to produce the final outputs Yr and the feedback outputs
ð2Þ s q h
Fl, with an efficiency of E0 = r = 1 ur Y r0 þ l = 1 cl F l0 = g = 1 wg Z g0 : The
model proposed by Liang et al. (2011) to measure the efficiency of this system
aims to maximize the average of the two division efficiencies, subject to the
constraints that describe the operations of the system. Under constant returns to
scale the model is
h s q
1 g = 1 wg Z g0 r = 1 ur Y r0 þ l = 1 cl F l0
E 0 = max: m q þ
i = 1 vi X i0 þ
2 h
l = 1 cl F l0 g = 1 wg Z g0
h m q
s:t: wg Z gj - vi X ij þ cl F lj ≤ 0, j = 1, . . . , n
g=1 i=1 l=1 ð11:1Þ
s q h
ur Y rj þ cl F lj - wg Z gj ≤ 0, j = 1, . . . , n
r=1 l=1 g=1
ur , vi , wg , cl ≥ ε, 8r, i, g, l
The objective function is nonlinear, and thus a nonlinear program solver is needed
to find the solution. At optimality, the division and system efficiencies are:
11.1 Feedback System 253
Fig. 11.1 Structure of the

feedback system
Xi Zg Yr
1 2
g=1,…, h r =1,…, s
i=1,…, m
Fl
l=1,…, q
h m q
ð1Þ
E0 = wg Z g0 = vi X i0 þ cl F l0
g=1 i=1 l=1
s q h
ð2Þ
E0 = ur Y r0 þ cl F l0 = wg Z g0
r=1 l=1 g=1
ð1Þ ð2Þ
E0 = E0 þ E0 =2
If a linear model is desired, then the fixed weight of 0.5 for the two division
efficiencies can be relaxed to allow for flexible weights, similar to that discussed in
Chen et al. (2009). Specifically, define the weight of a division as the proportion of
the aggregate input used by this division in that used by the two divisions, i.e.,
ωð1Þ = m
i = 1 vi X i0 þ
q
l = 1 cl F l0 =
m
i = 1 vi X i0 þ
h
g = 1 wg Z g0 þ
q
l = 1 cl F l0 and
ωð2Þ = h
g = 1 wg Z g0 =
m
i = 1 vi X i0 þ h
g = 1 wg Z g0 þ q
l = 1 cl F l0 : Since ω(1) + ω(2)
= 1 and ω , ω ≥ 0, the system efficiency is a weighted arithmetic average of the
(1) (2)
two division efficiencies. Each DMU in this case is seeking the most favorable
weight to yield the largest efficiency. Denote σ as the weighted average of the
division efficiencies, and this can be expressed as:
ð1Þ ð2Þ
σ = ωð1Þ × E0 þ ωð2Þ × E0
m q h
i = 1 vi X i0 þ l = 1 cl F l0 g = 1 wg Z g0
= × m q þ
i = 1 vi X i0 þ
m h q
i = 1 vi X i0 þ g = 1 wg Z g0 þ l = 1 cl F l0 l = 1 cl F l0
h s q
g = 1 wg Z g0 r = 1 ur Y r0 þ l = 1 cl F l0
m h q
× h
i = 1 vi X i0 þ g = 1 wg Z g0 þ l = 1 cl F l0 w Z
g = 1 g g0
s h q
r = 1 ur Y r0 þ g = 1 wg Z g0 þ l = 1 cl F l0
= m h q
i = 1 vi X i0 þ g = 1 wg Z g0 þ l = 1 cl F l0
The associated model is then:

s h q
r = 1 ur Y r0 þ g = 1 wg Z g0 þ l = 1 cl F l0
max: m
i = 1 vi X i0 þ
h
g = 1 wg Z g0 þ
q
l = 1 cl F l0
ð11:2Þ
s:t: Constraints of Modelð11:1Þ
Since this model is a linear fractional program, it can be linearized by applying

the variable substitution technique in Charnes and Cooper (1962) to find a solution.
Specifically, the denominator of the objective function is replaced with the inverse of
q
a variable t, m i = 1 vi X i0 þ
h
g = 1 wg Z g0 þ l = 1 cl F l0 = 1=t. Substituting ur = ur t,
vi = vi t, wg = wg t, and cl = cl t produces the following linear model:
s h q
max: ûr Y r0 þ ^ g Z g0 þ
w ^cl F l0
r=1 g=1 l=1
m h q
s:t: ^vi X i0 þ ^ g Z g0 þ
w ^cl F l0 = 1
i=1 g=1 l=1
h m q
^ g Z gj -
w ^vi X ij þ ^cl F lj ≤ 0, j = 1, . . . , n
g=1 i=1 l=1
s q h
ûr Y rj þ ^cl F lj - ^ g Z gj ≤ 0,
w j = 1, . . . , n
r=1 l=1 g=1
ûr , ^vi , w
^ g , ^cl ≥ εt, 8r, i, g, l, t > 0
At optimality, the objective value is the system efficiency, and the two sets of
ð1Þ
constraints indicate that E 0 = hg = 1 wg Z g0 = m
i = 1 vi X i0 þ
q
l = 1 cl F l0 and
ð2Þ s q h
E0 = r = 1 ur Y r0 þ l = 1 cl F l0 Þ= g = 1 wg Z g0 :
Both Models (11.1) and (11.2) are the efficiency aggregation model in additive
form, which defines the system efficiency as a weighted average of the division
efficiencies. The difference between them is that the former applies a set of
pre-defined weights to calculate the system efficiency, while the other uses a set of
data-reflected weights. Whether one should use pre-assigned or data-reflected
weights depends on the purpose of the efficiency measurement. However, the latter
has the advantage that the associated model is implicitly linear.
A closer examination of the two division efficiencies reveals that the numerator of
the efficiency of Division 1 is the same as the denominator of that of Division 2, such
that they will cancel out if one adopts the multiplicative form of efficiency aggre-
gation to define the system efficiency. The product form produces a system effi-
ciency which is implicitly linear and in a simpler form. Denote π as the product of the
two division efficiencies, we then have
h s q
ð1Þ ð2Þ g = 1 wg Z g0 r = 1 ur Y r0 þ l = 1 cl F l0
π = E0 × E0 = m q ×
i = 1 vi X i0 þ
h
l = 1 cl F l0 g = 1 wg Z g0
s q
r = 1 ur Y r0 þ l = 1 cl F l0
= m
q
vi X i0 þ l = 1 cl F l0
i=1
with the associated model of:

s q
r = 1 ur Y r0 þ l = 1 cl F l0
max: m q
i = 1 vi X i0 þ l = 1 cl F l0 ð11:3Þ
s:t: Constraints of Model ð11:1Þ
Again, this model is a linear fractional program, which can be linearized by

applying the variable substitution technique from Charnes and Cooper (1962).
Compared with Model (11.2), this model has a simpler objective function, in that
both the numerator and denominator do not have the terms corresponding to the
intermediate products.
As discussed in the preceding chapter, there are two concepts for measuring the
system efficiency: efficiency decomposition and aggregation. All three models
discussed thus far are based on the concept of efficiency aggregation. If the concept
of efficiency decomposition is to be used, then the objective function must be the
system efficiency, expressed as the ratio of the aggregate exogenous output to the
aggregate exogenous input. The efficiency decomposition model for the feedback
system is then:
s
r = 1 ur Y r0
E 0 = max: m
i = 1 vi X i0
h m q
s:t: wg Z gj - vi X ij þ cl F lj ≤ 0, j = 1, . . . , n
g=1 i=1 l=1 ð11:4Þ
s q h
ur Y rj þ cl F lj - wg Z gj ≤ 0, j = 1, . . . , n
r=1 l=1 g=1
ur , vi , wg , cl ≥ ε, 8r, i, g, l
Since the numerator and denominator of the objective function do not have
common variables, the fractional objective function can be linearized by assigning
1 to the denominator to be a constraint and leaving the numerator as the new
objective function.
Regarding the decomposition of system efficiency, one idea is to consider this
system as a variation of the basic two-stage system and to expect that its system
efficiency can be decomposed into division efficiencies similar to that of the basic
two-stage system. In the basic two-stage system, the system efficiency can be
decomposed into the product of the division efficiencies. Starting with this relation-
ship, we multiply the two division efficiencies together to get
h s q
ð1Þ ð2Þ g = 1 wg Z g0 r = 1 ur Y r0 þ l = 1 cl F l0
π = E0 × E0 = m q ×
i = 1 vi X i0 þ
h
l = 1 cl F l0 w Z
g = 1 g g0
ð11:5Þ
s q
r = 1 ur Y r0 þ l = 1 cl F l0 E þρ
= = 0
m
i = 1 vi X i0 þ
q
l = 1 cl F l0
1þρ
where the last fractional is obtained by dividing both the numerator and denominator
q
by m i = 1 vi X i0 and denoting ρ = l = 1 cl F l0 =
m
i = 1 vi X i0 . Since ρ is greater than or
equal to zero, this fractional is greater than or equal to the system efficiency E0.
Rearranging terms, Expression (11.5) becomes:
ð1Þ ð2Þ
E 0 = πð1 þ ρÞ - ρ = E 0 × E 0 ð1 þ ρ Þ - ρ ð11:6Þ
indicating that the system efficiency is the product of the two division efficiencies
adjusted by a factor ρ, the ratio of the aggregate feedback to the aggregate input.
Expression (11.6) implies that the system efficiency E0 will equal the product of
the division efficiencies π if π = 1, i.e., the two divisions are efficient, or ρ = 0, i.e.,
there is no feedback. This is reasonable, because, in the first case, the system will of
course be efficient if the two component divisions are efficient, and, in the second
case, the system becomes the basic two-stage one if there is no feedback, which has
ð1Þ ð2Þ
the relationship of E0 = E0 × E 0 obtained for the basic two-stage system in the
preceding chapter. Another property of this system, according to Expression (11.5),
is that the system efficiency E0 is less than or equal to the product of the two division
efficiencies π, and the larger ρ is the greater the difference between E0 and π. This is
also reasonable, because a small ρ implies that the aggregate feedback is relatively
small as compared to the aggregate exogenous input, which indicates that the system
is close to a basic two-stage system.
Model (11.4) can be linearized to formulate its dual as follows:
min: θ
n
ð1Þ
s:t: λj X ij ≤ θX i0 , i = 1, . . . , m
j=1
n n
ð1Þ ð2Þ
λj Z gj ≥ λj Z gj , g = 1, . . . , h
j=1 j=1
n
ð1Þ
n
ð2Þ
ð11:7Þ
λj F lj ≤ λj F lj ,
j=1 j=1
n
ð2Þ
λj Y rj ≥ Y r0 , r = 1, . . . , s, l = 1, . . . , q
j=1
ðkÞ
λj ≥ 0, j = 1, . . . , n, k = 1, 2
Note that the non-Archimedean number ε has been omitted for simplicity of
expression. This model shows the target for every factor, which can help the DMU to
ðk Þ
achieve efficient status. Let λj be a set of optimal solutions, the targets are:
n
^ i0 = ð1Þ
X λj X ij , i = 1, . . . , m
j=1
ðoutÞ n ðinÞ n
ð1Þ ð2Þ
Z^ g0 = λj Z gj , Z^ g0 = λj Z gj , g = 1, . . . , h
j=1 j=1
ðinÞ n n
^ l0
F = λj
ð1Þ
F lj , ^ ðoutÞ
F l0 =
ð2Þ
λj F lj , l = 1, . . . , q
j=1 j=1
n
ð2Þ
Y^ r0 = λj Y rj , r = 1, . . . , s,
j=1
ðoutÞ ðinÞ
The intermediate products have two targets, Z g0 and Z g0 , where the former is
the target for Division 1 to produce and the latter is that for Division 2 to consume.
ðoutÞ
Due to the second constraint of Model (11.7), Z g0 will be greater than or equal to
ðinÞ
Z g0 . In other words, the amount produced must be greater than or equal to that
ðinÞ ðoutÞ
consumed. The feedback also has two targets, F l0 and F l0 , where the former is the
target for Division 1 to consume and the latter is that for Division 2 to produce. The
amount to be consumed cannot exceed that to be produced, and this is ensured by the
third constraint of Model (11.7). Allowing for a difference between the quantities
supplied and consumed implies that strong disposability is assumed.
11.2 Independent Efficiency Measures
The general two-stage system discussed in this chapter has the structure shown in
ð1Þ
Fig. 11.2, where Division 1 applies the exogenous inputs X i , i = 1, . . ., m(1), to
produce the final outputs Y ðr1Þ , r = 1, . . ., s , and intermediate products Zg, g = 1,
(1)
ð2Þ
. . ., h and Division 2 applies the exogenous inputs X i , i = m(1) + 1, . . ., m, and
intermediate products Zg, g = 1, . . ., h, to produce the final outputs Y ðr2Þ ,
r = s(1) + 1, . . ., s. Since the inputs and outputs of the two divisions are different,
the superscript (k) for the inputs and outputs can actually be discarded. However, it is
still used for better identifying the division of concern.
The most straightforward way to measure the efficiency of this system is to treat
the two divisions as two independent systems and measure their efficiencies sepa-
rately by two DEA models. Under constant returns to scale, the models in ratio
form are:
h sð1Þ ð1Þ
ð1Þ g = 1 wg Z g0 þ r = 1 ur Y r0
E0 = max: mð1Þ ð1Þ
i = 1 vi X i0
h sð1Þ mð1Þ ð11:8aÞ
ð1Þ ð1Þ
s:t: wg Z gj þ ur Y rj - vi X ij ≤ 0, j = 1, . . . , n
g=1 r=1 i=1
ur , vi , wg ≥ ε, 8r, i, g
Fig. 11.2 Structure of the X i(2) , i = m(1) + 1,..., m

general two-stage system
X i(1) Zg Yr(2)
1 2
i = 1, ..., m (1) g=1,…, h r = s (1) +1, ..., s
Yr(1) , r = 1,..., s (1)

11.2 Independent Efficiency Measures 259
s ð2Þ
ð2Þ r = sð1Þ þ1 ur Y r0
E0 = max: m ð2Þ h
i = mð1Þ þ1 vi X i0 þ g = 1 wg Z g0
s m h
ð2Þ ð2Þ
s:t: ur Y rj - vi X ij þ wg Z gj ≤ 0, j = 1, . . . , n
r = sð1Þ þ1 i = mð1Þ þ1 g=1
ur , vi , wg ≥ ε, 8r, i, g
ð11:8bÞ
ð1Þ ð2Þ
After the division efficiencies E 0 and E0 are calculated from the above models,
the system efficiency can be defined and calculated, for example, as the average of
the division efficiencies. This type of measurement ignores the interactions between
the divisions. However, it is simple, and it gives a rough picture of how each DMU
has performed as compared to others.
Suppose there are five DMUs using one input X1 for Division 1 to produce one
final output Y1 and one intermediate product Z, and using the intermediate product
Z and one input X2 for Division 2 to produce the final output Y2, with the structure
shown in Fig. 11.3. Columns two to six of Table 11.1 show the data for the five
DMUs, and Fig. 11.4 shows the production of this system. The right half is the
production of Division 1, where the input X1 is used to produce the output Y1 and
intermediate product Z, and the production frontier is represented by the product
transformation curves of X1 = 1 and X1 = 2. The left half is the production of
Division 2, where the input X2 and intermediate product Z are used to produce the
Fig. 11.3 Structure of the X2

example system
Z
X1 1 2 Y2
Y1
Table 11.1 Data and independent efficiency measures of the example system
DMU X1 Y1 Z X2 Y2 ð1Þ ð2Þ ð1Þ ð2Þ
E0 E0 E0 þ E0 =2
A 1 1 0.5 4 1 0.5385 1 0.7692
B 2 1 5 1 1 1 1 1
C 1 1.5 2 6 2 1 0.9545 0.9773
D 2 3 3 3 2 0.9231 1 0.9615
E 2 4 1 4 1 1 0.8077 0.9038
Z
Division 2 Division 1
Y2 = 2
6
Y2 = 1 X1 = 2
B (2) B (1)
4
D̂ (1)
D (1)
D (2) X1 = 1
C (2) C (1)
2
(2)
Ĉ
E (2) Â(1)
E (1)
A (2) Ê (2) A (1)
X2 6 4 2 O 2 4 Y1
Fig. 11.4 Production frontiers of the example system
output Y2, and the production frontier is represented by the isoquants of Y2 = 1 and
Y2 = 2. The efficiencies of the five DMUs for the two divisions can be measured via
Models (11.8a) and (11.8b), with the results shown in columns seven and eight of
Table 11.1. Suppose the two divisions are considered as equally important. The last
column of Table 11.1 shows the system efficiencies of the five DMUs, which are the
averages of the two division efficiencies.
For Division 1, DMUs B, C, and E are efficient, which is confirmed by Fig. 11.4,
in that these three DMUs lie on the product transformation curves. For Division
2, DMUs A, B, and D have a perfect efficiency score of one, which is also confirmed
by Fig. 11.4, in that these DMUs lie on the isoquants. There is only one DMU, B,
which is efficient in both stages. Figure 11.4 also shows the target for the inefficient
DMUs, as denoted by the name of the DMU, with a caret on top of it. The average
division efficiencies in the last column of Table 11.1 indicate that the rankings of the
five DMUs are B, C, D, E, and A.
The merit of the ratio-form efficiency measurement model is that it is able to measure
the system and division efficiencies at the same time. There are two concepts of this
form used in measuring the efficiency of network systems: efficiency decomposition
and aggregation. They will be discussed separately in this section. In efficiency

aggregation there is an idea of treating the two divisions as the players in a two-
person game, which is unique to two-division systems. This section will start with
this game approach, followed by discussions of efficiency aggregation and
decomposition.
11.3.1 Game Approach
In studying the efficiency of supply chains, Liang et al. (2006) treated the seller
(Division 1) and buyer (Division 2) as two players in a game. They can either operate
independently or in a cooperative manner. If it is the former, then it is assumed that
one of the players is the leader, who makes a decision first, and the other is the
follower, who makes a decision after noting what the leader has decided. If it is the
latter, then the two players will work together to find the best decision for both of
them. The payoffs in this case are the efficiencies evaluated for the two divisions,
and the strategies are the multipliers selected by the two divisions to measure
efficiencies. In a supply chain the seller usually does not have final outputs Y ðr1Þ , as
discussed in Liang et al. (2006). However, to make the discussion more general, we
assume Division 1 also produces final outputs.
Suppose Divisions 1 and 2 do not cooperate, and Division 1 is the leader,
although it should be noted that the discussion when Division 2 is the leader
would be very similar. In the former case Division 1 tries to get the highest efficiency
measure for itself, without considering what the efficiency of Division 2 will
be. Division 1 will thus use the Independent Model (11.8a) to measure its efficiency.
Suppose a value of E(1) is obtained. After noticing the decision made by Division
1, which brings an efficiency measure of E(1) to that division, Division 2 then applies
the Independent Model (11.8b) to measure its own efficiency, under the condition
that the efficiency of Division 1 must be maintained at the value of E(1). The
associated model is:
s ð2Þ
ð2Þ r = sð1Þ þ1 ur Y r0
E0 = max: m ð2Þ h
i= mð1Þ þ1 vi X i0 þ g= 1 wg Z g0
sð1Þ h mð1Þ
ð1Þ ð1Þ
s:t: ur Y r0 þ wg Z g0 = Eð1Þ vi X i0
r=1 g=1 i=1
sð1Þ h mð1Þ ð11:9Þ

ð1Þ ð1Þ
ur Y rj þ wg Z gj - vi X ij ≤ 0, j= 1, ..., n
r =1 g=1 i=1
s m h
ð2Þ ð2Þ
ur Y rj - vi X ij þ wg Z gj ≤ 0, j= 1, ..., n
r =sð1Þ þ1 i = mð1Þ þ1 g=1
ur , vi , wg ≥ ε, 8r, i,g
The first constraint requires that no matter what strategy Division 2 is going to
use, the efficiency of Division 1 thus measured must be maintained at the value of
ð2Þ
E(1). The objective value E 0 is then the efficiency of Division 2, as a follower. By
applying Model (11.9) to the data in Table 11.1, the efficiency of Division 2 is the
same as that measured independently via Model (11.8b) for every DMU, as shown in
Table 11.1.
If the two divisions decide to cooperate with each other, then they will select a set
of multipliers that will produce the largest mutual benefit. In Liang et al. (2006) this
mutual benefit is expressed as the average of the two division efficiencies. The
associated model is:
sð1Þ ð1Þ h ð2Þ

r = 1 ur Y r0 þ
s
1 g = 1 wg Z g0 r = sð1Þ þ1 ur Y r0
E 0 = max: ð1Þ ð1Þ
þ ð2Þ
2 m h
i = mð1Þ þ1 vi X i0 þ
m
i = 1 vi X i0 g = 1 wg Z g0
sð1Þ h mð1Þ
ð1Þ ð1Þ
s:t: ur Y rj þ wg Z gj - vi X ij ≤ 0, j = 1, . . . , n
r=1 g=1 i=1
s m h
ð2Þ ð2Þ
ur Y rj - vi X ij þ wg Z gj ≤ 0, j = 1, . . . , n
ur , vi , wg ≥ ε, 8r, i, g
ð11:10Þ
This model is nonlinear. At optimality, the two fractionals in the objective

function are the efficiencies of the two divisions, and their average, which is the
objective value, is the efficiency of the system. For the data in Table 11.1, the
efficiencies of the two divisions measured from Model (11.10) are the same as those
measured independently via Models (11.8a) and (11.8b).
The efficiency of a division is its aggregate output divided by its aggregate input, as
ð1Þ ð2Þ
seen in E 0 and E 0 in Models (11.8a) and (11.8b) for the two divisions. The
concept of efficiency aggregation is based on defining the system efficiency as an
aggregation of the division efficiencies. There are two forms of aggregation that
appear in the literature: additive and multiplicative.
For the additive form, the system efficiency is defined as a weighted average of
the division efficiencies. The average efficiency used in the cooperative game model
(11.10) is an example of this form. Since this form is nonlinear, one idea is to
carefully select the weight in calculating the weighted average so that the model can
become (implicitly) linear. Suppose we define the weight associated with a division
to be the proportion of the aggregate input consumed by this division in that
consumed by both divisions. The weighted average of the two division efficiencies
becomes:
ð1Þ ð2Þ
ωð1Þ × E0 þ ωð2Þ × E0
m ð 1Þ ð1Þ sð1Þ ð1Þ h
i = 1 vi X i0 r = 1 ur Y r0 þ g = 1 wg Z g0
= m h
× m ð 1Þ ð1Þ
i = 1 vi X i0 þ g = 1 wg Z g0 i = 1 vi X i0
m ð2Þ h ð2Þ
i = mð1Þ þ1 vi X i0 þ g = 1 wg Z g0
s
r = sð1Þ þ1 ur Y r0
þ m h
× ð2Þ
i = 1 vi X i0 þ
m h
g = 1 wg Z g0 i = mð1Þ þ1 vi X i0 þ g = 1 wg Z g0
s h
r = 1 ur Y r0 þ g = 1 wg Z g0
= m h
i = 1 vi X i0 þ g = 1 wg Z g0

where X i0 = X i0 þ X i0 and Y r0 = Y r0 þ Y r0 . The associated model is then:
s h
r = 1 ur Y r0 þ g = 1 wg Z g0
E 0 = max: m h
i = 1 vi X i0 þ g = 1 wg Z g0
sð1Þ h mð1Þ
ð1Þ ð1Þ
r=1 g=1 i=1
s m h
ð2Þ ð2Þ
ur , vi , wg ≥ ε, 8r, i, g
ð11:11Þ
This model is a linear fractional program, which can be linearized by applying the
variable substitution technique in Charnes and Cooper (1962). The two constraint
ð1Þ sð1Þ ð1Þ

sets indicate that the efficiencies of the two divisions are E 0 = r = 1 ur Y r0 +
h mð1Þ ð1Þ ð2Þ s ð2Þ m ð2Þ h
g = 1 wg Z g0 Þ / i = 1 vi X i0 and E 0 = r = sð1Þ þ1 ur Y r0 = i = mð1Þ þ1 vi X i0 þ g=1
wg Z g0 Þ, and the objective value is the system efficiency E0.
For the multiplicative form, the system efficiency is defined as the product of the
division efficiencies, and the associated model is:
sð1Þ ð1Þ h ð2Þ

r = 1 ur Y r0 þ
s
g = 1 wg Z g0 r = sð1Þ þ1 ur Y r0
E0 = max: ð1Þ ð1Þ
× m ð2Þ h
m
i = 1 vi X i0 g = 1 wg Z g0
ð11:12Þ
This model is obviously nonlinear, and a nonlinear programming solver must be

used to find a solution. At optimality, the two fractionals in the objective function are
the two division efficiencies, and their product, which is the objective value, is the
system efficiency.
The concept of efficiency decomposition is based on measuring the efficiency of the

system, which is expressed as the ratio of the aggregate exogenous output to the
aggregate exogenous input, taking into account the operations of the divisions.
Specifically, the model is:
s
r = 1 ur Y r0
E 0 = max: m
i = 1 vi X i0 ð11:13Þ
This model can be linearized by assigning 1 to the denominator of the objective

function as a constraint and leaving the numerator as the new objective function. At
optimality, the two constraint sets indicate that the efficiencies of the two divisions
ð1Þ s ð 1Þ ð1Þ h m ð 1Þ ð1Þ ð2Þ s ð2Þ
are E0 = r = 1 ur Y r0 þ g = 1 wg Z g0 = i = 1 vi X i0 and E 0 = r = sð1Þ þ1 ur Y r0 =
m ð2Þ h
i = mð1Þ þ1 vi X i0 þ g = 1 wg Z g0 , and the objective function shows that the sys-
tem efficiency is E 0 ¼ sr¼1 ur Y r0 =
m
i¼1 vi X i0 : The question is how to decompose
the system efficiency into a function of the division efficiencies, from which the
division that has a greater effect on the efficiency of the system can be identified. An
improvement in this division will thus increase the efficiency of the overall system to
a greater extent. There are two ways to accomplish this.
Table 11.2 Additive efficiency decomposition for the example system

ð1Þ ð2Þ
DMU E0 (ω(1)) E0 (ω(2)) ρ σ E0
A 0.5385 (0.7558) 1 (0.2442) 0.0617 0.6512 0.6296
B 1 (0.3143) 1 (0.6857) 0.4 1 1
C 1 (0.5963) 0.9545 (0.4037) 0.2247 0.9817 0.9775
D 0.9231 (0.3514) 1 (0.6486) 0.0882 0.9730 0.9706
E 1 (0.9993) 0.8077 (0.0007) 0.0003 0.9999 0.9999
The first is to assume that the system efficiency is related to the weighted average
of the division efficiencies and try to derive the relationship between them. Let σ
denote the weighted average of the division efficiencies, where the weight is the
proportion of the aggregate input consumed by the division in that consumed by the
two divisions. Based on the derivation of Model (11.11), we have σ= ωð1Þ ×
ð1Þ ð2Þ
E 0 þ ωð2Þ × E 0 = s
r = 1 ur Y r0 þ h
g = 1 wg Z g0 = m
i = 1 vi X i0 þ h
g = 1 wg Z g0 :
m
Dividing both the numerator and denominator by i = 1 vi X i0 obtains:
ð1Þ ð2Þ E0 þ ρ
σ = ωð1Þ × E 0 þ ωð2Þ × E 0 = ð11:14Þ
1þρ
where ρ = hg = 1 wg Z g0 = m
i = 1 vi X i0 is the ratio of the aggregate intermediate product
to the total aggregate exogenous input. Rearranging terms, E0 can be expressed in
terms of σ and ρ as:
ð1Þ ð2Þ
E 0 = σð1 þ ρÞ - ρ = ωð1Þ × E 0 þ ωð2Þ × E 0 ð1 þ ρ Þ - ρ ð11:15Þ
This expression shows that the system efficiency is equal to the weighted average
of the division efficiencies adjusted by a factor ρ. There are two cases to be noted.
One is that when the two divisions are efficient, i.e. σ = 1, the system will also be
efficient, with E0 = 1. The other is that if there are no intermediate products, i.e.,
ρ = 0, then the system efficiency E0 will equal the weighted average of the division
efficiencies σ. If one of these two cases does not apply, then Expression (11.14)
implies that E0 will be smaller than σ. Moreover, the larger ρ is, the greater the
difference between the system efficiency E0 and the weighted average of the division
efficiencies σ.
For the five DMUs in Table 11.1, their efficiencies are calculated from Model
(11.13), with the results shown in Table 11.2. It is noted that the two division
efficiencies happen to be the same as those measured from the independent model,
as shown in Table 11.1. The weights selected by the DMUs to calculate the weighted
average are different. Using DMU A to explain the relationship between the system
ð1Þ ð2Þ
efficiency E0 and the two division efficiencies E 0 and E 0 expressed in (11.15), the
I II
X i(1) Zg Zg Yr( 2 )
1 2
i = 1, ..., m (1) Y r
(1)
X i
( 2) r = s (1) +1, ..., s
Xi Yr
r = 1, ..., s (1) i = m(1) +1, ..., m
i=1,..., m r =1,..., s
X i( 2) X i( 2) Yr(1) Yr(1)
3 4
i = m(1) +1, ..., m i = m(1) + 1, ..., m r = 1, ..., s (1) r = 1, ..., s (1)
Fig. 11.5 A series-parallel transformation of the general two-stage system
weighted average of the division efficiencies is 0.7558×0.5385+

0.2442×1 = 0.6512, which is equal to σ. The value of σ(1 + ρ) - ρ = 0.6512×(1+
0.0617) - 0.0617 = 0.6296, which is the value of E0. It is also noted that the system
efficiency is less than the weighted average of the division efficiencies, unless the
system is efficient (e.g., DMU B).
The second way of decomposing the system efficiency is to transform the
structure of the general two-stage system to one of the basic two-stage system.
The idea, as suggested by Kao (2009), is to introduce dummy divisions which
produce the outputs as its inputs. By introducing dummy division 3, which con-
ð2Þ ð2Þ
sumes X i to produce X i , and dummy division 4, which consumes Y ðr1Þ to produce
Y ðr1Þ , the structure of the general two-stage system in Fig. 11.2 can be transformed
into the one shown in Fig. 11.5, which is a basic two-stage structure composed of
two subsystems labeled as I and II, and each subsystem has two divisions connected
in parallel. The two real divisions are represented by squares and the two dummy
ones are represented by circles.
The system in Fig. 11.5 has four divisions, with the same exogenous inputs Xi,
i = 1, . . ., m, and exogenous outputs Yr, r = 1, . . ., s. The complete relational model
that describes this system is:
s
r = 1 ur Y r0
E 0 = max: m
i = 1 vi X i0
s:t: System constraints :
s m
ur Y rj - vi X ij ≤ 0, j= 1, ..., n
r=1 i=1
Subsystem constraints :
sð1Þ h m m
ð1Þ ð2Þ
ðIÞ ur Y rj þ wg Z gj þ vi X ij - vi X ij ≤0, j= 1, ..., n
r =1 g=1 i= mð1Þ þ1 i= 1
s sð1Þ h m
ð1Þ ð2Þ
ðIIÞ ur Y rj - ur Y rj þ wg Z gj þ vi X ij ≤ 0, j= 1, ... ,n
r=1 r=1 g= 1 i =mð1Þ þ1
ð11:16Þ
sð1Þ h mð1Þ
ð1Þ ð1Þ
ð1Þ ur Y rj þ wg Z gj - vi X ij ≤ 0, j = 1, . . . , n
r=1 g=1 i=1
s m h
ð2Þ ð2Þ
ð2Þ ur Y rj - vi X ij þ wg Z gj ≤ 0, j = 1, . . . , n
m m
ð2Þ ð2Þ
ð3Þ vi X ij - vi X ij ≤ 0, j = 1, . . . , n
i = mð1Þ þ1 i = mð1Þ þ1
sð1Þ sð1Þ
ð1Þ ð1Þ
ð4Þ ur Y rj - ur Y rj ≤ 0, j = 1, . . . , n
r=1 r=1
ur , vi , wg ≥ ε, r = 1, . . . , s, i = 1, . . . , m, g = 1, . . . , h
The sum of the two subsystem constraints (I) and (II) is equal to the system
constraint for each DMU. The system constraints are thus redundant and can be
deleted. The sum of the division constraints (1) and (3) is equal to subsystem
constraint (I) for each DMU, and the sum of the division constraints (2) and (4) is
equal to subsystem constraint (II), which makes the subsystem constraints (I) and
(II) redundant. Finally, the division constraints (3) and (4) are identities, which also
are redundant, and can be deleted. Deleting the redundant constraints leaves only
division constraints (1) and (2) as the active ones, which indicates that Model (11.16)
is the same as Model (11.13), the one corresponding to the original structure.
Denote E I0 and E II0 as the efficiencies of Subsystems I and II, respectively. Based
on Model (11.16), we have
s ð 1Þ h m m
ð1Þ ð2Þ
E I0 = ur Y r0 þ wg Z g0 þ vi X i0 = vi X i0
r=1 g=1 i = mð1Þ þ1 i=1
s sð1Þ h m
ð1Þ ð2Þ
E II0 = ur Y r0 = ur Y r0 þ wg Z g0 þ vi X i0
r=1 r=1 g=1 i = mð1Þ þ1
The product of the two subsystem efficiencies is just the efficiency of the system,
E I0 × EII0 = E 0 , fulfilling the relationship of the series structure.
In Subsystem I the efficiency of Division 1, as before, is the ratio of the aggregate
ð1Þ s ð 1Þ ð1Þ h mð1Þ ð1Þ
output to aggregate input, E0 = r = 1 ur Y r0 þ g = 1 wg Z g0 = i = 1 vi X i0 : The
efficiency of the dummy division 3 is also the ratio of its aggregate output to its
ð3Þ ð2Þ ð2Þ
aggregate input, E 0 = m i = mð1Þ þ1 vi X i0 =
m
i = mð1Þ þ1 vi X i0 , which is one. If we aggre-
gate the efficiencies of Divisions 1 and 3 together additively by applying the weight
of the proportion of the aggregate input associated with each division in that
associated with the two divisions, we have
ð1Þ ð3Þ
ωð1Þ × E 0 þ ωð3Þ × E 0
mð1Þ ð1Þ sð1Þ ð1Þ h ð2Þ
r = 1 ur Y r0 þ
m
i = 1 vi X i0 g = 1 wg Z g0 i = mð1Þ þ1 vi X i0
= m × mð1Þ ð1Þ
þ m ×
i = 1 vi X i0 i = 1 vi X i0 i = 1 vi X i0
m ð2Þ
i = mð1Þ þ1 vi X i0
m ð2Þ
sð1Þ ð1Þ h m ð2Þ
r = 1 ur Y r0 þ g = 1 wg Z g0 þ i = mð1Þ þ1 vi X i0
= m
i = 1 vi X i0
= E I0
which is equal to the efficiency of Subsystem I. By the same token, the weighted
average of the efficiencies of Divisions 2 and 4 is equal to the efficiency of
Subsystem II, derived as:
ð2Þ ð4Þ
ωð2Þ ×E 0 þ ωð4Þ ×E 0
m ð2Þ h ð2Þ
i =mð1Þ þ1 vi X i0 þ
s
g = 1 wg Z g0 r = sð1Þ þ1 ur Y r0
= sð1Þ ð1Þ ð2Þ
× m ð2Þ h
h m
r =1 ur Y r0 þ g = 1 wg Z g0 þ i= mð1Þ þ1 vi X i0 g = 1 wg Z g0
sð1Þ ð1Þ sð1Þ ð1Þ
r = 1 ur Y r0 r = 1 ur Y r0
þ sð1Þ ð1Þ h m ð2Þ
× sð1Þ ð1Þ
r =1 ur Y r0 þ g = 1 wg Z g0 þ i = mð1Þ þ1 vi X i0 r = 1 ur Y r0
s
r = 1 ur Y r0
= sð1Þ ð1Þ h m ð2Þ
= EII0
r =1 ur Y r0 þ g = 1 wg Z g0 þ i= mð1Þ þ1 vi X i0
Combining all these together, we have

E 0 = E I0 × E II0 = ωð1Þ × E0 þ ωð3Þ × E0 × ωð2Þ × E 0 þ ωð4Þ × E 0
ð1Þ ð2Þ
= ωð1Þ × E0 þ ωð3Þ Þ × ðωð2Þ × E 0 þ ωð4Þ
ð11:17Þ
where ω(3) = 1 - ω(1) and ω(4) = 1 - ω(2). This is another way of decomposing the
system efficiency. When ω(3) and ω(4) are small, which implies that the exogenous
inputs of Division 2 and the exogenous outputs of Division 1 are relatively small, the
system is close to a series one, with the system efficiency E0 close to the product of
ð1Þ ð2Þ
the division efficiencies E 0 × E0 .
We next use the data in Table 11.1 to explain the relationship in (11.17). By
applying Model (11.16) the division, subsystem, system efficiencies, and weights are
ð1Þ
calculated as shown in Table 11.3. For DMU A, ωð1Þ × E0 þ ωð3Þ = 0:8025 ×
0.5385 + (1 - 0.8025 ) = 0.6296, which is equal to E I0 , as required by Expression
ð2Þ
(11.17). Similarly, ωð2Þ × E 0 þ ωð4Þ = 0:4118 × 1 þ ð1 - 0:4118 Þ = 1, which is
II
equal to E 0 . Finally, the product of the two subsystem efficiencies is
0.6296×1 = 0.6296, which is equal to the system efficiency E0. Note that the weights
associated with the divisions in this decomposition, ω(1) and ω(2), shown in
Table 11.3 Multiplicative efficiency decomposition for the example system

ð1Þ ð2Þ
DMU E0 (ω(1)) E0 (ω(2)) EI0 E II0 E0
A 0.5385 (0.8025) 1 (0.4118) 0.6296 1 0.6296
B 1 (0.4400) 1 (0.9600) 1 1 1
C 1 (0.7303) 0.9545 (0.4944) 1 0.9775 0.9775
D 0.9231 (0.3824) 1 (0.7273) 0.9706 1 0.9706
E 1 (0.9996) 0.8077 (0.0007) 1 0.9999 0.9999
Table 11.3, are different from those of the other decomposition, shown in Table 11.2,
ð1Þ ð2Þ
although the E0 , E 0 , and E0 are the same.
The dual of the ratio-form model has the envelopment form, which is equivalent to
the distance function form of the DEA model for measuring efficiencies. Its merit is
that it is able to show the targets for inefficient DMUs to become efficient. Distance
functions can be defined for the system, divisions, and the input–output factors. It
can also be directional. All these types of distance function have been applied to the
general two-stage system to measure efficiencies. Since the input–output factors
approach essentially produces slacks-based efficiency measures (referring to the case
discussed in the preceding chapter), it will not be discussed again here.
Färe and Whittaker (1995) formulated perhaps the first distance function model to
measure the system efficiency of the general two-stage system. The basic idea is
convexity of the production possibility set and strong disposability of the factors.
Under constant returns to scale, the model corresponding to the system with the
structure shown in Fig. 11.2 from the input side is:
min:θ
n
ð1Þ ð1Þ ð1Þ
s:t: λj X ij ≤ θX i0 , i = 1, . . . , mð1Þ
j=1
n
ð2Þ ð2Þ ð2Þ
λj X ij ≤ θX i0 , i = mð1Þ þ 1, . . . , m
j=1
n
ð1Þ
λj Z gj ≥ Z g0 , g = 1, . . . , h
j=1
n
ð2Þ
ð11:18Þ
λj Z gj ≤ Z g0 , g = 1, . . . , h
j=1
n
ð1Þ ð1Þ ð1Þ
λj Y rj ≥ Y r0 , r = 1, . . . , sð1Þ
j=1
n
ð2Þ ð2Þ ð2Þ
λj Y rj ≥ Y r0 , r = sð1Þ þ 1, . . . , s
j=1
ðkÞ
λj ≥ 0, j = 1, . . . , n, k = 1, 2
Similar to the ratio-form models, the non-Archimedean number ε has been

ignored for simplicity of expression. If the two divisions have variable returns to
ðk Þ
scale technologies, then the constraints nj = 1 λj = 1, k = 1, 2 are added.
Model (11.18) is not able to show the efficiencies of the two divisions. However,
it is able to show the targets of the factors, a function not shared with the ratio-form
models. Specifically, the targets are:
n
^ ð1Þ
X i0 =
ð1Þ ð1Þ
λj X ij , i = 1, . . . , mð1Þ
j=1
n
^ ð2Þ
X i0 =
ð2Þ ð2Þ
λj X ij , i = mð1Þ þ 1, . . . , m
j=1
ðoutÞ n ðinÞ n
ð1Þ ð2Þ
Z^ g0 = λj Z gj , Z^ g0 = λj Z gj , g = 1, . . . , h ð11:19Þ
j=1 j=1
ð1Þ n
ð1Þ ð1Þ
Y^ r0 = λj Y rj , r = 1, . . . , sð1Þ
j=1
ð2Þ n
ð2Þ ð2Þ
Y^ r0 = λj Y rj , r = sð1Þ þ 1, . . . , s
j=1
It is interesting to note that the target for the intermediate products perceived by
Division 1 does not necessarily need to be the same as that perceived by Division
ðoutÞ
2. However, the former, Z g0 , must be greater than, or at least equal to, the latter,
ðinÞ
Z g0 , to be feasible.
The distance function for the general two-stage system can also be formulated
from the output side, which is:
max:φ
n
ð1Þ ð1Þ ð1Þ
s:t: λj X ij ≤ X i0 , i = 1, . . . , mð1Þ
j=1
n
ð2Þ ð2Þ ð2Þ
λj X ij ≤ X i0 , i = mð1Þ þ 1, . . . , m
j=1
n
ð1Þ
λj Z gj ≥ Z g0 , g = 1, . . . , h
j=1
n ð11:20Þ
ð2Þ
λj Z gj ≤ Z g0 , g = 1, . . . , h
j=1
n
ð1Þ ð1Þ ð1Þ
λj Y rj ≥ φY r0 , r = 1, . . . , sð1Þ
j=1
n
ð2Þ ð2Þ ð2Þ
λj Y rj ≥ φY r0 , r = sð1Þ þ 1, . . . , s
j=1
ðkÞ
λj ≥ 0, j = 1, . . . , n, k = 1, 2
In this form, the system efficiency is the inverse of the objective function 1/φ. It
can be shown, from the ratio-form model, that the system efficiency measured from
the input side, θ, is equal to that measured from the output side, 1/φ. Whether one
should use the input or output model depends on the aims of a particular project.
The ratio-form model (11.13) discussed in efficiency decomposition can be
expressed in the multiplier form, which has a dual of the following form:
Table 11.4 System distance function measures for the example system
ðoutÞ ðinÞ
DMU E0 X1 X2 Z Z Y1 Y2
A 1 1 4 1.8333 0.5 1 1
(0.6296) (0.6296) (2.5185) (1.0926) (1.0926) (1) (1)
B 1 2 1 5 5 1 1
(1) (2) (1) (5) (5) (1) (1)
C 0.9995 1 5.5 2 2 1.5 2
(0.9775) (0.9775) (5.8652) (1.8539) (1.8539) (1.5) (2)
D 1 2 3 4 3 3 2
(0.9706) (1.9412) (2.9118) (3.6176) (3.6176) (3) (2)
E 0.9999 2 2.75 1 1 4 1
(0.9999) (2) (2.75) (1) (1) (4) (1)
min:θ
n
ð1Þ ð1Þ ð1Þ
s:t: λj X ij ≤ θX i0 , i = 1, . . . , mð1Þ
j=1
n
ð2Þ ð2Þ ð2Þ
λj X ij ≤ θX i0 , i = mð1Þ þ 1, . . . , m
j=1
n n
ð1Þ ð2Þ
λj Z gj ≥ λj Z gj , g = 1, . . . , h ð11:21Þ
j=1 j=1
n
ð1Þ ð1Þ ð1Þ
λj Y rj ≥ Y r0 , r = 1, . . . , sð1Þ
j=1
n
ð2Þ ð2Þ ð2Þ
λj Y rj ≥ Y r0 , r = sð1Þ þ 1, . . . , s
j=1
ðkÞ
λj ≥ 0, j = 1, . . . , n, k = 1, 2
This model is the same as Model (11.18), except that the third constraint for the
intermediate products is a combination of the third and fourth constraints of Model
(11.18). Since this model is less stringent than Model (11.18), the efficiency mea-
sured from it is less than or equal to that measured from Model (11.18). The targets
for the factors have the same forms as those shown in Expression (11.19).
Applying Models (11.18) and (11.21) to the data contained in Table 11.1 pro-
duces the results shown in Table 11.4, where the numbers in the second row of each
DMU in parentheses are measured from Model (11.21). There are several points to
be noted. First, as expected, the system efficiency measured from Model (11.21) is
less than or equal to that measured from Model (11.18) for every DMU. Second, the
system efficiency measured from Model (11.21) is the same as that measured from
Model (11.13), due to the primal-dual relationship between these two models. Third,
since Models (11.18) and (11.21) are input-oriented, which seek the smallest amount
of inputs to become efficient while maintaining the outputs at the current level, the
targets of the outputs are the same as their current level. Fourth, the targets for the
intermediate product Z may be different for the two divisions, as indicated by DMUs
A and D calculated from Model (11.18). All these findings are common to other
examples.
The input-oriented model seeks the smallest amount of inputs while maintaining the
outputs at the current level to become efficient. Conversely, the output-oriented
model looks for the largest amount of outputs while maintaining the inputs at the
current level for achieving efficiency. One idea is to combine these two models to
reconcile the inputs to be reduced and the outputs to be expanded.
The input-oriented model aims to minimize the input contraction parameter θ, and
the output-oriented model aims to maximize the output expansion parameter φ. One
way to combine these two models is to minimize the sum of positive θ and negative
φ, and thus obtain the following model:
min:θ - φ
n
ð1Þ ð1Þ ð1Þ
s:t: λj X ij ≤ θX i0 , i = 1, . . . , mð1Þ
j=1
n
ð2Þ ð2Þ ð2Þ
λj X ij ≤ θX i0 , i = mð1Þ þ 1, . . . , m
j=1
n n
ð1Þ ð2Þ
λj Z gj ≥ λj Z gj , g = 1, . . . , h
j=1 j=1 ð11:22Þ
n
ð1Þ ð1Þ ð1Þ
λj Y rj ≥ φY r0 , r = 1, . . . , sð1Þ
j=1
n
ð2Þ ð2Þ ð2Þ
λj Y rj ≥ φY r0 , r = sð1Þ þ 1, . . . , s
j=1
θ ≤ 1, φ≥1
ðkÞ
λj ≥ 0, j = 1, . . . , n, k = 1, 2
Here the idea of the relational model, as used in Model (11.21), is adopted in
formulating the constraints corresponding to the intermediate products. The dual of
this model is as follows:
Table 11.5 Division distance function measures for the example system. X
ð1Þ ð2Þ
DMU θ φ X1 X2 Z Y1 Y2 E0 E0 E0
A 1 1.5882 1 4 1.7353 1.5882 1.5882 0.5385 1 0.6296
B 1 1 2 1 5 1 1 1 1 1
C 1 1.0230 1 6 1.8966 1.5345 2.0460 1 0.9545 0.9775
D 1 1.0303 2 3 3.7273 3.0909 2.0606 0.9231 1 0.9706
E 1 1 2 2.75 1 4 1 1 0.8077 0.9999
max: - α þ β
m
s:t: vi X i0 - α = 1
i=1
s
ur Y r0 - β = 1
r=1
sð1Þ h mð1Þ ð11:23Þ
ð1Þ ð1Þ
ur Y rj þ wg Z gj - vi X ij ≤ 0, j = 1, . . . , n
r=1 g=1 i=1
s m h
ð2Þ ð2Þ
α, β, ur , vi , wg ≥ 0, 8r, i, g
Substituting the constraints of m i = 1 vi X i0 - α = 1 and

s
r = 1 ur Y r0 - β = 1 into
s m
the objective function obtains - α þ β = r = 1 ur Y r0 - i = 1 vi X i0 , which indicates
that Model (11.22) is an additive model. From the constraints of Model (11.23), we
ð1Þ s ð 1Þ ð1Þ h m ð 1Þ ð1Þ ð2Þ s ð2Þ
have E 0 = r = 1 ur Y r0 þ g = 1 wg Z g0 = i = 1 vi X i0 , E 0 = r = sð1Þ þ1 ur Y r0 =
m ð2Þ h s m
i = mð1Þ þ1 vi X i0 þ g = 1 wg Z g0 Þ, and E 0 = r = 1 ur Y r0 = i = 1 vi X i0 : To examine
whether this model can do what is expected, we apply Models (11.22) and (11.23) to
the data in Table 11.1, and thus obtain the results shown in the first eight and last
three columns, respectively, of Table 11.5. The system efficiency in the last column
shows that DMU B is strongly efficient, DMU E is weakly efficient, and the
remaining three are inefficient.
It is noted that all DMUs have a perfect efficiency measure for the inputs, even for
the four inefficient ones, A, C, D, and E, indicating that the current level of inputs
need not be reduced. This is also confirmed by the target values X 1 and X 2 in
columns four and five, respectively, where only the X 2 of DMU E is smaller than its
current value due to the condition of weak efficiency. This result is the same as that
of the basic two-stage system discussed in Sect. 10.3.2 of the preceding chapter. The
three inefficient DMUs, A, C, and D, are reflected by the output expansion parameter
φ, with a value greater than one. The target values Y 1 and Y 2 in columns seven and
eight, respectively, for the inefficient DMUs are their current values multiplied by φ.
Model (11.22) is thus not able to reconcile the effect of excessive consumption of the
inputs and insufficient production of the outputs. The target values for the
intermediate product Z from the viewpoints of Divisions 1 and 2 are the same, and
thus only one is shown (in column six). The last three columns show that the division
and system efficiencies measured from Model (11.23) are the same as those mea-
sured from Model (11.13) and shown in Table 11.3.
Another way of combining the minimization of θ and the maximization of φ is to
minimize the ratio of θ to φ, to yield the following model:
min:θ=φ
ð11:24Þ
variable substitution technique in Charnes and Cooper (1962). For the same reason
discussed in Model (11.22), the effect of excessive consumption of the inputs and
insufficient production of the outputs cannot be reconciled by this model.
In order to obtain the division and system efficiencies one must formulate the dual
of this model and then make the related calculations based on the corresponding
constraints. Kao (2017) found that the target identified by the envelopment model is
not the one used by the ratio model to calculate division efficiencies. Based on the
primal-dual relationship of the two models, the envelopment model was
reformulated to obtain the target and measure the division efficiencies at the
same time.
11.4.3 Directional Distance Parameter
The input-oriented model reduces the excessive inputs and the output-oriented one
expands the insufficient outputs, maintaining the other factors at the current level in
measuring efficiencies. The directional distance function approach is to move along
a pre-specified direction toward the frontier, such that the excessive inputs will be
reduced and the insufficient outputs will be expanded.
Let d = (-f, g) be the direction. The directional distance function approach is to
find the parameter η that the inputs can be reduced and the outputs can be expanded
in this proportion. The associated model is:
Table 11.6 Targets obtained from the directional distance function approach
ðoutÞ ðinÞ
DMU η X1 X2 Z Z Y1 Y2
A 0.2273 0.7727 3.0909 1.3409 1.3409 1.2273 1.2273
B 0 2 1 5 5 1 1
C 0.0114 0.9886 5.9318 1.875 1.875 1.5170 2.0227
D 0.0149 1.9701 2.9552 3.6716 3.6716 3.0448 2.0299
E 0 2 4 1 0.5 4 1
max:η
n
s:t: λj X ij ≤ X i0 - η f i , i = 1, . . . , mð1Þ
j=1
n
λj X ij ≤ X i0 - ηf i , i = mð1Þ þ 1, . . . , m
j=1
n n
ð1Þ ð2Þ
λj Z gj ≥ λj Z gj , g = 1, . . . , h ð11:25Þ
j=1 j=1
n
λj Y rj ≥ Y r0 þ ηgr , r = 1, . . . , sð1Þ
j=1
n
λj Y rj ≥ Y r0 þ ηgr , r = sð1Þ þ 1, . . . , s
j=1
ðkÞ
λj ≥ 0, j = 1, . . . , n, k = 1, 2
Consider the example in Table 11.1. Suppose the direction of d = (-X0, Y0) is
used. The targets for the inputs, outputs, and intermediate products for the five
DMUs, as obtained from Model (11.25), are shown in Table 11.6. Recall that an
efficient DMU has a value of zero for η, and larger values of η indicate conditions of
lower efficiency. There are two DMUs, B and E, which have η = 0, where the former
is strongly efficient, and the latter is weakly efficient. As expected, all inefficient
DMUs reduce some amount of the inputs and increase some amount of the outputs to
become efficient.
The directional distance function approach is especially useful for cases with
undesirable factors. Different distance parameters can also be assigned to different
factors, if this is desired.
The radial efficiency measurement either fixes the outputs at the current level to
measure the excessive inputs consumed, as in the input-oriented model, or fixes the
inputs at the current level to measure the insufficient outputs that should have been
produced, as in the output-oriented model. The slacks-based model takes both the
excessive inputs consumed and the insufficient outputs produced into account in
measuring efficiencies. For general two-stage systems with the structure shown in
Fig. 11.2, the slack variables for measuring efficiencies are generated from the
constraints corresponding to the production possibility set defined in the
following form:
n
λj X ij þ si = X i0 , i = 1, . . . , mð1Þ
j=1
n
λj X ij þ si = X i0 , i = mð1Þ þ 1, . . . , m
j=1
n
ð1Þ
λj Z gj - ^sþ
g = Z g0 g = 1, . . . , h
j=1
n
ð2Þ ð11:26Þ
λj Z gj þ ^sg- = Z g0 , g = 1, . . . , h
j=1
n
λj Y rj - sr = Y r0 , r = 1, . . . , sð1Þ
j=1
n
λj Y rj - sr = Y r0 , r = sð1Þ þ 1, . . . , s
j=1
λ, s ≥ 0
This set of constraints assumes a constant returns to scale technology. If Division

ðk Þ
k has a variable returns to scale technology, then the constraint nj = 1 λj = 1 is
added. The efficiencies of the two divisions, as defined in Tone (2001), are:
ð1Þ -
1 mð1Þ si
1-
ð1Þ
mð1Þ i=1
X i0
ð1Þ
E0 =
1 sð1Þ
ð1Þþ
sr h ^sþ
g
1þ þ
ðsð1Þ þ hÞ r=1
Y r0
ð1Þ g = 1 Z g0
ð11:27Þ
1 m
ð2Þ -
si h ^sg-
1- þ
ð2Þ
ðm - mð1Þ þ hÞ i = mð1Þ þ1
X i0
ð2Þ g = 1 Z g0
E0 = ð2Þþ
1 s sr
1þ
s - sð1Þ r = sð1Þ þ1
Y r0
ð2Þ
Tone and Tsutsui (2010) proposed using a set of pre-specified weights ω(k) to
combine the slack variables related to the two divisions in the following form to
measure the system efficiency. The related model is:
1
ð1Þ-
mð1Þ si 1 m
ð2Þ -
si h ^sg-
ωð1Þ 1- þ ωð2Þ 1- þ
mð1Þ i =1 ð1Þ
X i0 ðm- mð1Þ þ hÞ i= mð1Þ þ1
X i0
ð2Þ g= 1 Z g0
min:
1
ð1Þþ
sð1Þ sr h ^sþ
g 1 s
ð2Þþ
sr
ωð1Þ 1 þ þ þ ωð2Þ 1 þ
ðsð1Þ þ hÞ r =1 ð1Þ
Y r0 g =1 Z g0 s -sð1Þ r = sð1Þ þ1
Y r0
ð2Þ
s:t: Expressionð11:26Þ
n n
ð1Þ ð2Þ
λj Z gj = λj Z gj , g =1, ...,h
j =1 j =1
ð11:28Þ
Note that a set of constraints that requires the intermediate products to be equal
between the two divisions are added to keep continuity. Recall that this set of
constraints is in the greater than or equal to form in the relational model to allow
for disposition. Model (11.28) is nonlinear and can be linearized by applying the
variable substituting technique in Charnes and Cooper (1962). At optimality, the
objective value is the system efficiency, and the two division efficiencies are as
defined in Expression (11.27). The method of handling non-discretionary and
undesirable factors discussed in Chap. 6 can be incorporated into this model.
Kao (2014b) proposed the idea of using the flexible weights of
þ
sð1Þ sðr1Þþ h sg
1 þ ðsð1Þ1þhÞ r = 1 Y ð 1Þ þ g = 1 Z g0
ωð1Þ =
r0
þ
sð1Þ sðr1Þþ h sg s ð2Þþ
1 þ ðsð1Þ1þhÞ r = 1 Y ð 1Þ þ g = 1 Z g0 þ 1 þ s -1sð1Þ sr
r = sð1Þ þ1 Y ð2Þ
r0 r0
ð2Þþ
s
1 þ s -1sð1Þ sr
ωð 2 Þ = þ
r0
sð1Þ srð1Þþ h sg s ð2Þþ

1 þ ðsð1Þ1þhÞ r = 1 Y ð 1Þ þ g = 1 Z g0 þ 1 þ s -1sð1Þ sr
r0 r0
to obtain the weighted average of:
ð1Þ ð2Þ
ωð1Þ × E0 þ ωð2Þ × E0
ð1Þ - ð2Þ -
mð1Þ si m si h ^sg-
1- 1
mð1Þ i = 1 X ð1Þ þ 1- 1
ðm - mð1Þ þhÞ i = mð1Þ þ1 X ð2Þ þ g = 1 Z g0
=
i0 i0
sð1Þ sð1Þþ h ^sþ s ð2Þþ

1 þ ðsð1Þ1þhÞ r
r = 1 Y ð1Þ þ g
g = 1 Z g0 þ 1 þ s -1sð1Þ sr
r0 r0
as the system efficiency. The associated model is:
1
ð1Þ-
mð1Þ si 1 m si
ð2Þ-
h ^sg-
1- þ 1- þ
mð1Þ i=1 ð1Þ
X i0 ðm-mð1Þ þhÞ i=mð1Þ þ1
X i0
ð2Þ g=1 Z g0
min:
1 sð1Þ
ð1Þþ
sr h ^sþ
g 1 s sr
ð2Þþ
1þ þ þ 1þ
ðsð1Þ þhÞ r=1
Y r0
ð1Þ g=1 Z g0 s-sð1Þ r=sð1Þ þ1
Y r0
ð2Þ
s:t: Expressionð11:26Þ
n n
ð1Þ ð2Þ
λj Z gj = λj Z gj , g=1, ...,h
j=1 j=1
ð11:29Þ
The merit of this model is that the system efficiency is a weighted average of the
division efficiencies, where the weights are the most favorable ones for obtaining the
largest possible system efficiency. It is also noted that if ω(1) and ω(2) in Model
(11.28) are assigned a value of 0.5, then Model (11.28) boils down to (11.29),
indicating that the efficiencies measured using the approach from Tone and Tsutsui
(2010) are the same as those obtained using the method in Kao (2014b).
Applying Model (11.29) to the data in Table 11.1 obtains the results shown in
Table 11.7. It is interesting to note that in this example the two division efficiencies
ð1Þ ð2Þ
E 0 and E0 are the same as those measured from the independent radial measures,
Table 11.7 Slacks-based ð1Þ ð2Þ

DMU E0 E0 (ω(1)) E0 (ω(2))
efficiency measures for the
example system A 0.7692 0.5385 (0.5) 1 (0.5)
B 1 1 (0.5) 1 (0.5)
C 0.9767 1 (0.4884) 0.9545 (0.5116)
D 0.9615 0.9231 (0.5) 1 (0.5)
E 0.8936 1 (0.4468) 0.8077 (0.5532)
11.6 Shared Input 281
shown in Table 11.1. The system efficiency E0 is the weighted average of the two
ð1Þ ð2Þ
division efficiencies E 0 and E 0 .
11.6 Shared Input
In network production systems there are cases in which an input is shared by several
divisions for their production. For example, a professor can spend some of his/her
time at work on teaching and the remaining on conducting research. If the times
allocated to the two tasks by the professor are known, then the conventional DEA
models discussed in the preceding sections are able to measure the related efficien-
cies. However, if the time allocation is not clear, then it falls into the category of a
shared-input system.
Consider the basic two-stage system, with the inputs being shared by the two
divisions, as depicted in Fig. 11.6. Let αi and (1-αi) denote the proportions of input
Xi allocated to Divisions 1 and 2, respectively. If the ith input is solely used by
Division 1, then we have αi = 1. In contrast, if input Xi is solely used by Division
2, then we have αi = 0. Since the percentage of the input that is allocated to a division
is not clear, the proportion that will yield the greatest system efficiency is conven-
tionally sought, so that the results are more acceptable to the DMUs being evaluated.
In order to exclude unreasonable results, suitable lower and upper bounds for αi are
usually imposed.
When αi and (1-αi) of the ith input are used by Divisions 1 and 2, respectively,
the corresponding amounts are αiXi and (1-αi)Xi. Let Li and Ui be the lower and
upper bounds, respectively, of αi. The multiplier-form model under constant returns
to scale can be formulated as:
Fig. 11.6 The basic two-

stage system with shared
input
1 - αi
Xi Zg Yr
1 2
i=1,…, m g =1,…, h r =1,…, s
s
E 0 = max: ur Y r0
r=1
m
s:t: vi X i0 = 1
i=1
h m
wg Z gj - vi ðαi X ij Þ ≤ 0, j = 1, . . . , n
g=1 i=1
s m h
ur Y rj - vi ð1 - αi ÞX ij þ wg Z gj ≤ 0, j = 1, . . . , n
r=1 i=1 g=1
Li ≤ αi ≤ U i , i = 1, . . . , m
ur , vi , wg ≥ ε, 8r, i, g
ð11:30Þ
Due to the nonlinear terms of viαi, this model is nonlinear. One way to linearize
this model is to substitute vi = vi αi to obtain the following:
s
E0 = max: ur Y r0
r=1
m
s:t: vi X i0 = 1
i=1
h m
wg Z gj - ^vi X ij ≤ 0, j = 1, . . . , n
g=1 i=1
s m h
ur Y rj - ðvi - ^vi ÞX ij þ wg Z gj ≤ 0, j = 1, . . . , n
r=1 i=1 g=1
vi Li ≤ ^vi ≤ vi U i , i = 1, . . . , m
ur , vi , wg ≥ ε, 8r, i, g
At optimality, αi = vi =vi reflects the proportion of input Xi that the DMU would
like to allocate to Division 1 to yield the measured efficiencies. The objective value
ð1Þ
is the efficiency of the system, and the two division efficiencies are E 0 =
ð2Þ
h
g=1 wg Z g0 = m i = 1 vi X i0 and E0 =
s
r = 1 ur Y r0 =
m
i=1 vi - vi X i0 þ h
g = 1 wg Z g0 :
Consider five DMUs using a portion of the input X for Division 1 to produce the
intermediate product Z, and the remaining of X, together with all of Z, for Division
2 to produce the output Y. The data shown in columns two to four of Table 11.8 are
the same as those contained in Table 10.1. Suppose it is not clear what percentage of
the input X is used by Divisions 1 and 2 for the production. However, it is known that
neither division will consume more than 75% of it. By applying Model (11.30), the
results shown in the last five columns of Table 11.8 are obtained. The α value
Table 11.8 Data and efficiency measures for the example of shared input
ð1Þ ð2Þ
DMU X Z Y α E0 E0 E0 (α = 1)
A 2 1 0.5 0.25 1/2 0.2714 0.2375 (0.2)
B 4 2 1 0.25 1/2 0.2714 0.2375 (0.2)
C 3 3 2 0.25 1 0.6333 0.6333 (0.5333)
D 5 4 5 0.25 4/5 1 0.95 (0.8)
E 6 5 5.5 0.25 5/6 0.9087 0.8708 (0.7333)
Fig. 11.7 The basic two- X 1(1) X 1(2)

stage system with the first
input allowed to be
reallocated
Xi Zg Yr
1 2
i=2,..., m g =1,…, h r =1,…, s
reaches the lower bound of 0.25 for all five DMUs, indicating that using less X for
Division 1 will be more beneficial to the system efficiency. For the original case of
the basic two-stage system, where all of the input X is used by Division 1 (with
α = 1), the system efficiency is smaller, as indicated by the values in parentheses in
the last column of Table 11.8. In this regard, sharing some of the input with Division
2 increases the overall efficiency.
In the above discussion the key issue is that it is not clear how the input is shared
by the two divisions, while the aim is to achieve the highest efficiency that can be
obtained under the most favorable conditions. Another situation is that the input has
been allocated to the two divisions, but it is not clear whether the allocation is the
best, in terms of the efficiency score, that can be achieved. Färe et al. (1997) thus
proposed the following model to find the optimal allocation for the input.
Consider a basic two-stage system, with the first input allowed to be reallocated to
ð1Þ ð2Þ
the two divisions, as illustrated in Fig. 11.7. Currently, X 1j and X 1j are used by the
ð1Þ ð2Þ
two divisions of DMU j. Suppose the total amount of X 1j = X 1j þ X 1j is to be
reallocated to the two divisions with the objective of increasing the efficiency of the
ð1Þ
system as much as possible. The idea in Färe et al. (1997) is to find the amounts x1
ð2Þ
and x1 that should be used by the two divisions via the following output-oriented
model:
max:φ
n
ð1Þ ð1Þ ð1Þ
s:t: λj X 1j ≤ x1
j=1
n
ð1Þ ð1Þ
λj X ij ≤ X i0 , i = 2, . . . , m
j=1
n n
ð1Þ ð2Þ
λj Z gj ≥ λj Z gj , g = 1, . . . , h
j=1 j=1
ð11:31Þ
n
ð2Þ ð2Þ ð2Þ
λj X 1j ≤ x1
j=1
n
ð2Þ
λj Y rj ≥ φY r0 , r = 1, . . . , s
j=1
ð1Þ ð2Þ
x1 þ x1 = X 10
ð1Þ ð2Þ ðkÞ
x1 , x1 , λj ≥ 0, j = 1, . . . n, k = 1, 2
ð1Þ ð2Þ
At optimality, x1 and x1 are the amounts of input X1 to be allocated to the two
divisions. The objective value φ* shows how much the outputs can be expanded with
the new allocation.
Consider an example of five DMUs, where input X1 is shared by the two
ð1Þ
divisions. Division 1 uses X 1 and X2 to produce one intermediate product Z, and
ð2Þ
Division 2 uses X 1 and the intermediate product Z to produce one final output Y,
with the data shown in Table 11.9. Under the current situation that the first input is
ð1Þ ð2Þ
already assigned to the two divisions for production, the variables x1 and x1 in
ð1Þ ð2Þ
Model (11.31) are replaced with the current values of X 10 and X 10 , respectively, to
find the factor φ that shows by how much the outputs can be expanded. The results
are shown in column seven of Table 11.9 under the heading φ. Of the five DMUs,
only E is efficient, indicating that its output cannot be expanded any further, while
DMU A has the largest space for expansion, if it operates efficiently. If the input X1 is
allowed to be reallocated, then Model (11.31) is applied to find the expansion factor
and how X1 should be allocated, with the results shown in the last two columns under
ð1Þ ð2Þ
the headings of φ* and (x1 , x1 ). The expansion factor φ* is greater than that
ð1Þ ð2Þ
measured using the current values of X 10 and X 10 for every DMU, indicating that
Table 11.9 Data and efficiency measures for the resource allocation example

DMU X1 X1 , X1 X2 Z Y φ φ* x1 , x1
A 3 (1, 2) 2 2 1 3.3333 3.6 (0.6, 2.4)
B 4 (2, 2) 3 1 3 1.1111 1.6 (0.8, 3.2)
C 5 (2, 3) 3 1 2 2.5 3 (1.0, 4.0)
D 6 (3, 3) 4 2 4 1.25 1.8 (1.2, 4.8)
E 7 (4, 3) 4 3 5 1 1.68 (1.4, 5.6)
the output can be increased further if the input X10 is allocated in a better manner. It is
interesting to note that although DMU E is efficient under the current allocation, the
output can still be expanded by 1.68 times if the allocation of X1 is changed from the
current amount of (4, 3) to the best amount of (1.4, 5.6).
Different ideas and assumptions have been used in modeling network systems. In
the resource allocation case, for example, it is noted that the same resource X1 has
been treated as two different ones in formulating the constraints. This assumption is
necessary in order to find the optimal allocation. If they are treated as the same, then
the constraint will be nj = 1 λj X 1j þ nj = 1 λj X 1j ≤ x1 þ x1 = X 10 , which is not
helpful in finding the optimal allocation.
The first work on network DEA was carried out by Charnes et al. (1986), which
examined army recruitment, and in this the system was divided into two divisions,
awareness creation and contract establishment, and the efficiency of each process
was measured independently. Other applications using independent efficiency mea-
sures include the following works. Soteriou and Zenios (1999) investigated the
operational and marketing measures of quality of 144 branches of a commercial
bank in Cyprus. Abad et al. (2004) applied a two-stage DEA to link financial data to
firm value, and two independent BCC models were applied to 30 stocks in the
Spanish manufacturing industry in the years 1991–1996. Simon et al. (2011) ana-
lyzed the productivity growth of 34 Spanish university libraries using a Malmquist
productivity index. The library production system was composed of two stages:
internal production and service delivery.
For the ratio-form efficiency measures, Fang and Zhang (2008) proposed a model
similar to that of Liang et al. (2006), which allows some inputs to be allocated
between the two stages. Bichou (2011) used the same model to study the operations
of container ports, which were viewed as supply chains between the gate and yard-
quay. Ten container terminals from various countries were examined, based on data
for 70 terminal-years. Premachandra et al. (2012) extended the additive model of
Chen et al. (2009) for the basic two-stage system to allow for exogenous inputs for
the second stage in studying the performance of 66 large mutual fund companies in
the USA. Guan and Chen (2012) and Guan and Zuo (2014) used the same model to
measure the innovation efficiency of 22 OECD and 35 countries, respectively. Lu
et al. (2016) evaluated the management and investment efficiencies of 36 Investment
Trust Corporations in Taiwan for the period 2007–2011 in another application. Yuan
et al. (2015) used a multiplicative efficiency aggregation model to measure the
drought hazard and vulnerability of 31 provinces, municipalities, and autonomous
regions in China during 2006–2011. Khalili-Damghani and Shahmir (2015) used the
same model to discuss the case of interval data. Amirteimoori (2013) studied a car
manufacturing problem, where some of the products were defective and thus needed
to be repaired. Wu and Birge (2012) proposed another type of ratio-form division

efficiency model to measure the system and process efficiencies, when the prices of
all outputs are known. This model can be used to analyze potential gains from the
merger of two serial-chain models, and a case of 36 branches from a large Canadian
bank undergoing a banking operations merger was used to illustrate the proposed
approach. Kao and Liu (2022) employed a simulation technique to measure the
deposit collection and utilization efficiencies of 22 commercial banks in Taiwan with
stochastic correlated data. The system efficiency can also be decomposed into
division efficiencies.
The distance function approach has also been widely discussed in the literature.
For example, Färe and Whittaker (1995) developed an output system parameter
model to measure 137 dairy farms in Lake States of the USA. Färe and Grosskopf
(1996) then proposed a similar model with the Malmquist productivity index also
being calculated. Yang et al. (2008) studied a problem of farrow-to-finish swine
production, where pollution was produced in the first division and a pollution
abatement technology was applied in the second, with the objective of maximizing
the output system distance parameter. The performance of 39 randomly selected
farms in Taiwan was measured using the proposed model. Chiu, Huang, and Ting
(2011b) extended the idea in Chen and Zhu (2004) for basic two-stage systems to
cases where the second division also consumes exogenous inputs to measure the
repair performance for stricken cultivated land and the agricultural efficiency of
31 regions in China. Chiu et al. (2012) used the same model to measure the
performance of R&D, as a value chain framework, of 21 high-tech businesses in
China. Chiu, Huang, and Ma (2011a) modified the model to allow for uncontrollable
inputs, undesirable intermediate products, and undesirable outputs to measure the
transit and economic efficiencies of 30 regions of China. Liu et al. (2015) measured
the performance of the national innovation system of 40 countries from the aspects
of knowledge production and commercialization. Xie et al. (2012) analyzed the
environmental efficiency, in addition to economic benefits, in relation to the efforts
of power systems to achieve sustainable development, where undesirable outputs
appeared in both the generation and grid (supply) divisions. The objective was to
minimize the weighted input division distance parameters, and 30 provincial admin-
istrative regions in China were used for illustration. Chen et al. (2012) developed a
factor distance measure model to examine the trade-offs between efficiency enhance-
ment and pollution abatement of 22 incineration plants in Taiwan. The system had
two divisions, waste treatment and energy generation, where the latter had both
desired and undesired outputs. Maghbouli et al. (2014) proposed cooperative and
non-cooperative game approaches to measure the efficiency based on distance
parameters. Hampf (2014) devised an environmental efficiency measure of the
envelopment form which can be decomposed into the product of the production
and abatement efficiencies and was applied to 23 power plants in the USA. Wu et al.
(2015) evaluated the efficiency of energy saving and emission of 30 provinces,
municipalities, and autonomous regions in China.
The distance function can also be directional. For example, Lozano et al. (2013)
separated the airport operations into aircraft movement and landing in studying the
performance of 39 Spanish airports, with undesirable outputs generated from the first
division. Taking the non-storable feature into account, Yu (2008) divided the
transportation service into two stages to study the production efficiency and service
effectiveness of 40 railways around the world. Yang (2009) used an output direc-
tional distance measure model to measure the production and environmental effi-
ciencies of farrow-to-finish pig production. In the production division, pigs were
produced as final outputs, with wastes as undesirable intermediate products that were
handled by a wastewater control division. The performance of 31 pig farms in
Taiwan was evaluated.
The slacks-based approach of Tone and Tsutsui (2009) serves as a good basis to
start with in this area. Lozano and Gutiérrez (2014) measured the production and
sales efficiencies of 16 European airlines. Song et al. (2015) examined changes in
production and environmental efficiencies among 20 listed petroleum enterprises in
China. Avkiran and McCrystal (2012), who used the ranges of the inputs and outputs
of all DMUs to adjust the slack variables and found that the conventional
observation-adjusted and their own range-adjusted approaches produced positive
and significant ranking correlations, is another interesting study on this topic.
Shared-input systems have attracted much attention in the literature on network
DEA. For example, Golany et al. (2006) proposed three models to measure the
efficiency of a two-stage system with shared inputs. The system allows each division
to acquire resources from the other in exchange for delivery of the appropriate
intermediate or final products. Löthgren and Tambour (1999) included customer
satisfaction in studying the performance of 31 Swedish pharmacies. The system
consists of the production and consumption divisions, and a firm-specific allocation
of input resources to the two divisions was allowed. Kao and Hwang (2010)
extended the relational model of Kao and Hwang (2008) for the basic two-stage
system to allow for the inputs to be shared with the second stage. Chen et al. (2006)
used the average of the two division efficiencies as the objective function, and Chen
et al. (2010) used a weighted average, to find the best distribution of IT-based
resources. Yu and Lee (2009) used the minimum of the ratio of the input parameter
to the output parameter as the objective function to measure the production and
marketing efficiencies of 58 international tourist hotels in Taiwan, all with shared
inputs. Yu and Chen (2011) applied the same model to measure the production and
consumption efficiencies of 15 domestic air routes of an airline in Taiwan. The game
approach has also been applied to systems with shared inputs. For example, Zha and
Liang (2010) used the product of the two process efficiencies as the profit to be
maximized, and 30 US commercial banks were used to illustrate this approach.
Finally, Li et al. (2012) used the model in Chen et al. (2006) to study the regional
R&D process of 30 provincial level regions in China.
References
Abad, C., Thore, S. A., & Laffarga, J. (2004). Fundamental analysis of stock by two-stage DEA.
Managerial and Decision Economics, 25, 231–241.
Amirteimoori, A. (2013). A DEA two-stage decision processes with shared resources. Central
European Journal of Operations Research, 21, 141–151.
Avkiran, N. K., & McCrystal, A. (2012). Sensitivity analysis of network DEA: NSBM versus
NRAM. Applied Mathematics and Computation, 218, 11226–11239.
Bichou, K. (2011). A two-stage supply chain DEA model for measuring container-terminal
efficiency. International Journal of Shipping and Transport Logistics, 3, 6–26.
Report CCS #532. Center for Cybernetic Studies, University of Texas-Austin.
Chen, P. C., Chang, C. C., Yu, M. M., & Hsu, S. H. (2012). Performance measurement for
incineration plants using multi-activity network data envelopment analysis: The case of Taiwan.
Journal of Environmental Management, 93, 95–103.
Chen, Y., Cook, W. D., Li, N., & Zhu, J. (2009). Additive efficiency decomposition in two-stage
Chen, Y., Du, J., Sherman, H. D., & Zhu, J. (2010). DEA model with shared resources and
efficiency decomposition. European Journal of Operational Research, 207, 339–349.
Chen, Y., Liang, L., Yang, F., & Zhu, J. (2006). Evaluation of information technology investment:
A data envelopment analysis approach. Computers & Operations Research, 33, 1368–1379.
Chen, Y., & Zhu, J. (2004). Measuring information technology’s indirect impact on firm perfor-
mance. Information Technology and Management, 5, 9–22.
Chiu, Y. H., Huang, C. W., & Chen, Y. C. (2012). The R&D value-chain efficiency measurement
for high-tech industries in China. Asia Pacific Journal of Management, 29, 989–1006.
Chiu, Y. H., Huang, C. W., & Ma, C. M. (2011a). Assessment of China transit and economic
efficiencies in a modified value-chains DEA model. European Journal of Operational
Research, 209, 95–103.
Chiu, Y. H., Huang, C. W., & Ting, C. T. (2011b). Measuring the repair performance for stricken
cultivated land and agricultural efficiency in China with a modified two-stage DEA model. Asia-
Pacific Journal of Operational Research, 28, 633–649.
Fang, L., & Zhang, C. Q. (2008). Resource allocation based on the DEA model. Journal of the
Operational Research Society, 59, 1136–1141.
Färe, R., Grabowski, R., Grosskopf, S., & Kraft, S. (1997). Efficiency of a fixed but allocatable
input: A non-parametric approach. Economics Letters, 56, 187–193.
Färe, R., & Grosskopf, S. (1996). Productivity and intermediate products: A frontier approach.
Economics Letters, 50, 65–70.
Färe, R., & Whittaker, G. (1995). An intermediate input model of dairy production using complex
survey data. Journal of Agricultural Economics, 46, 201–213.
Golany, B., Hackman, S. T., & Passy, U. (2006). An efficiency measurement framework for multi-
stage production systems. Annals of Operations Research, 145, 51–68.
Guan, J., & Chen, K. (2012). Modeling the relative efficiency of national innovation systems.
Research Policy, 41, 102–115.
Guan, J., & Zuo, K. (2014). A cross-country comparison of innovation efficiency. Scientometrics,
100, 541–575.
Hampf, B. (2014). Separating environmental efficiency into production and abatement efficiency: A
nonparametric model with application to US power plants. Journal of Productivity Analysis, 41,
457–473.
References 289

Kao, C. (2014a). Network data envelopment analysis: A review. European Journal of Operational
Kao, C. (2017). Efficiency measurement and frontier projection identification for general two-stage
systems in data envelopment analysis. European Journal of Operational Research, 261,
679–689.
Research, 185, 418–429.
Kao, C., & Liu, S. T. (2022). Stochastic efficiencies of network production systems with correlated
stochastic data: The case of Taiwanese commercial banks. Annals of Operations Research, 315,
1151–1174.
Khalili-Damghani, K., & Shahmir, Z. (2015). Uncertain network data envelopment analysis with
undesirable outputs to evaluate the efficiency of electricity power production and distribution
processes. Computers & Industrial Engineering, 88, 131–150.
Li, Y., Chen, Y., Liang, L., & Xie, J. H. (2012). DEA models for extended two-stage network
structures. Omega, 40, 611–618.
Liang, L., Li, Z. Q., & Cook, W. D. (2011). Data envelopment analysis efficiency in two-stage
networks with feedback. IIE Transactions, 43, 309–322.
Liang, L., Yang, F., Cook, W. D., & Zhu, J. (2006). DEA models for supply chain efficiency
evaluation. Annals of Operations Research, 145, 35–49.
Liu, J. S., Lu, W. M., & Ho, M. H. C. (2015). National characteristics: Innovation systems from the
process efficiency perspective. R&D Management, 45, 317–338.
Löthgren, M., & Tambour, M. (1999). Productivity and customer satisfaction in Swedish pharma-
cies: A DEA network model. European Journal of Operational Research, 115, 449–458.
Lozano, S., & Gutiérrez, E. (2014). A slacks-based network DEA efficiency analysis of European
airlines. Transportation Planning and Technology, 37, 623–637.
Lozano, S., Gutiérrez, E., & Moreno, P. (2013). Network DEA approach to airports performance
assessment considering undesirable outputs. Applied Mathematical Modelling, 37, 1665–1676.
Lu, W. M., Liu, J. S., Kweh, Q. L., & Wang, C. W. (2016). Exploring the benchmarks of the
Taiwanese investment trust corporations: Management and investment efficiency perspectives.
Maghbouli, M., Amirteimoori, A., & Kordrostami, S. (2014). Two-stage network structures with
undesirable outputs: A DEA based approach. Measurement, 48, 109–118.
Premachandra, I. M., Zhu, J., Watson, J., & Galagedera, D. U. A. (2012). Best-performing US
mutual fund families from 1993 to 2008: Evidence from a novel two-stage DEA model for
efficiency decomposition. Journal of Banking & Finance, 36, 3302–3317.
Simon, J., Simon, C., & Arias, A. (2011). Changes in productivity of Spanish university libraries.
Omega, 39, 578–588.
Song, M., Zhang, J., & Wang, S. (2015). Review of the network environmental efficiencies of listed
petroleum enterprises in China. Renewable & Sustainable Energy Reviews, 43, 65–71.
Soteriou, A., & Zenios, S. A. (1999). Operations, quality, and profitability in the provision of
banking services. Management Science, 45, 1221–1238.
Tone, K., & Tsutsui, M. (2009). Network DEA: A slacks-based approach. European Journal of
3–4.
Wu, D. D., & Birge, J. R. (2012). Serial chain merger evaluation model and application to mortgage
banking. Decision Sciences, 43, 5–35.
Wu, J., Lin, L. V., Sun, J., & Ji, X. (2015). A comprehensive analysis of China’s regional energy
saving and emission reduction efficiency: From production and treatment perspectives. Energy
Policy, 84, 166–176.
Xie, B. C., Fan, Y., & Qu, Q. Q. (2012). Does generation form influence environmental efficiency
performance? An analysis of China's power system. Applied Energy, 96, 261–271.
Yang, C. C. (2009). Productive efficiency, environmental efficiency and their determinants in
farrow-to-finish pig farming in Taiwan. Livestock Science, 126, 195–205.
Yang, C. C., Hsiao, C. K., & Yu, M. M. (2008). Technical efficiency and impact of environmental
regulations in farrow-to-finish swine production in Taiwan. Agricultural Economics, 39, 51–61.
Yu, M. M. (2008). Assessing the technical efficiency, service effectiveness, and technical effec-
tiveness of the world's railways through NDEA analysis. Transportation Research Part
A-Policy and Practice, 42, 1283–1294.
Yu, M. M., & Chen, P. C. (2011). Measuring air routes performance using a fractional network data
envelopment analysis model. Central European Journal of Operations Research, 19, 81–98.
Yu, M. M., & Lee, B. C. Y. (2009). Efficiency and effectiveness of service business: Evidence from
international tourist hotels in Taiwan. Tourism Management, 30, 571–580.
Yuan, X. C., Tang, B. J., & Wei, Y. M. (2015). China’s regional drought risk under climate change:
A two-stage process assessment approach. Natural Hazards, 76, 667–684.
Zha, Y., & Liang, L. (2010). Two-stage cooperation model with input freely distributed among the
stages. European Journal of Operational Research, 205, 332–338.
Chapter 12
Multi-Stage Systems
The preceding two chapters discussed systems with two divisions connected in
series. Intuitively, the two-stage system should be able to be extended to multiple
stages to suit more general cases. As a matter of fact, many real-world systems have
a multi-stage structure, with assembly lines as typical examples, where raw materials
go through a number of workstations to become the final products. The meaning of
multi-stage system in this context is rather vague, because a stage may have several
divisions connected in different structures. What it refers to in the conventional
network DEA is a system composed of a number of divisions connected in series,
with only one division in each stage. In this regard, the term series system may be
more appropriate, and in this chapter these two terms will be used interchangeably
when there is no ambiguity.
Several models have been proposed to measure the efficiency of general multi-
stage systems. When a model is proposed it is usually based on a structure, and three
types of structures are discussed in this chapter. The simplest type is the basic series
structure, where all the outputs of a division become the inputs of the succeeding
stage, and every division does not consume exogenous inputs, except the first one.
The second type is the general series structure, where each division may consume
exogenous inputs, and some of its outputs may become exogenous outputs, flowing
out of the system. The third type is the series structure with reversal links, where the
outputs of a division in a later stage may be sent back to a division in an earlier stage
to use. Since most related studies are concerned with systems with the general series
structure, this chapter will focus more on these and divide the discussion according
to the ways in which the efficiencies are measured, i.e., independent, ratio-form,
distance function, and slacks-based efficiency measures.
https://doi.org/10.1007/978-3-031-27593-7_12
292 12 Multi-Stage Systems
12.1 Basic Series Structure
Chap. 10 discussed the basic two-stage system, where the first stage consumes all the
exogenous inputs to produce the outputs which are all consumed by the second stage
to produce the final outputs. The outputs of the first stage are usually referred to as
intermediate products. This type of system can be generalized to multiple stages as
the basic series system, with the structure shown in Fig. 12.1.
In network analysis there are two basic structures, series and parallel, that are the
basis for studying systems with more complicated structures, and the basic series
structure is the former. As shown in Fig. 12.1, the first division of the system
consumes all the exogenous inputs Xi supplied from outside to produce the interme-
diate products Z ðg1Þ , which are then used by the second division to produce the
intermediate products Z ðg2Þ . This process is continued until the last division p,
where its outputs become the final outputs of the system, Yr. No divisions consume
exogenous inputs, except the first one, and no division produces exogenous outputs,
except the last one. Many models can be used to measure the efficiency of this type
of system. In this chapter we will introduce only two of them, product decomposition
and additive aggregation, because these have some interesting properties. Note that
we assume every division produces different intermediate products. Specifically,
Division k produces intermediate products Z ðgkÞ , g = h(k - 1) + 1, . . ., h(k), k = 1, . . .,
p - 1, where h(0) = 0 and h( p - 1) = h. In this case the superscript (k) in Z gðkÞ is
actually not needed. However, it is still used for better identifying the division of
concern.
The concept of efficiency decomposition is to measure the efficiency of a system

based on the exogenous inputs it uses and the exogenous outputs it produces, while
taking the operations of the component divisions into account. A relationship
between the system and division efficiencies is then sought, where the efficiency
of a division is measured based on the inputs it consumes and outputs it produces.
The characteristic of this method of efficiency measurement is that the efficiency of a
unit, no matter whether it is the system or a division, is always the aggregation of its
Xi Z g(1) Z g( k – 1) Z g(k ) Z g( p – 1) Yr
1 … k … p
i=1,..., m g = 1, ..., h (1) g = h ( k – 2) + 1, ..., h ( k –1) g = h (k –1) + 1, ..., h (k ) g = h ( p– 2) + 1, ..., h ( p–1) r =1,..., s
Fig. 12.1 The basic series structure

12.1 Basic Series Structure 293
outputs divided by the aggregation of its inputs. Based on this concept, the efficiency
decomposition model under constant returns to scale can be formulated as:
s
r = 1 ur Y r0
E0 = max: m
i = 1 vi X i0
s m
s:t: ur Y rj - vi X ij ≤ 0, j = 1, . . . , n
r=1 i=1
hð1Þ m
ð1Þ
wg Z gj - vi X ij ≤ 0, j = 1, . . . , n
g=1 i=1
hðkÞ hðk - 1Þ
ðkÞ ðk - 1Þ
wg Z gj - wg Z gj ≤ 0, k = 2, . . . , p - 1, j = 1, . . . , n
ðk - 1Þ ðk - 2Þ
g=h þ1 g=h þ1
s hðp - 1Þ
ðp - 1Þ
ur Y rj - wg Z gj ≤ 0, j = 1, . . . , n
r =1 g = hðp - 2Þ þ1
ur , vi , wg ≥ ε, 8r, i, g
ð12:1Þ
where the first set of constraints describes the operation of the system as a whole
unit, and the remaining sets describe the operations of the p divisions. Since the
constraint corresponding to the system (the one in the first constraint set) is just the
sum of the p constraints corresponding to the p divisions of the system, it is
redundant and can be deleted without altering the solution.
At optimality, the objective value is the system efficiency, and the division
efficiencies, based on the constraints of Model (12.1), are:
hð1Þ m
ð1Þ ð1 Þ
E0 = wg Z g0 = vi X i0
g=1 i=1
ðk Þ
E0
ðk Þ ðk-1Þ
h h
ðk Þ ðk-1Þ
= wg Z gj = wg Z gj , k=2, ...,p
g=hðk-1Þ þ1 g=hðk-2Þ þ1
s hðp-1Þ
ðpÞ ðp-1Þ
-1,E 0 = ur Y r0 = wg Z gj ð12:2Þ
r =1 g=hðp-2Þ þ1
Clearly, the product of the p division efficiencies is just the system efficiency:
ðk Þ
E 0 = pk = 1 E 0 . In other words, the system efficiency can be decomposed into the
product of the division efficiencies. In the basic two-stage system it was derived that
the system efficiency is the product of the two division efficiencies. The relationship
obtained here is just a generalization from two to multiple divisions.
Fig. 12.2 An example

system with the basic series Z1 Z2 Z3
structure X 1 2 3 4 Y
Table 12.1 Data and efficiencies measured from the efficiency decomposition model for the basic
series example
DMU X Z1 Z2 Z3 Y E0 E0 E0 E0 E0
A 2 2 3 3 3 1/8 1/2 1 1/2 1/2
B 1 2 2 2 3 1/4 1 2/3 1/2 3/4
C 3 2 1 2 2 1/18 1/3 1/3 1 1/2
D 2 3 2 3 2 1/12 3/4 4/9 3/4 1/3
E 1 1 1 1 2 1/6 1/2 2/3 1/2 1
As an example, consider five DMUs with a basic series structure shown in

Fig. 12.2. There are four divisions, where the first division uses the exogenous
input X to produce the intermediate product Z1. The second division applies the
intermediate product Z1 to produce another intermediate product Z2, and the third
division in turn applies the intermediate product Z2 to produce the intermediate
product Z3. Finally, the fourth division uses the intermediate product Z3 to produce
the final output Y. Columns two to six of Table 12.1 show the data for the five
DMUs. By applying Model (12.1), the system efficiency of each DMU is solved, as
shown in column seven. Based on Eq. (12.2), the efficiencies of the four divisions
are calculated, with the results shown in the last four columns of Table 12.1. It is
clear that, for each DMU, the system efficiency is the product of the four-division
efficiencies.
In this simple example the optimal solution is unique. For more complicated
systems where more than one intermediate product is produced by each division,
there may exist multiple solutions. In this case the system efficiency remains the
same. The division efficiencies, however, may vary for different solutions of the
multipliers, making them incomparable among different DMUs. A common basis
for measuring the efficiency of each division is thus necessary to make them
comparable, and an approach similar to that discussed in Model (10.3) can be
applied. Suppose the efficiencies of Division 1 for all DMUs are to be compared.
After the system efficiency E0 is measured from Model (12.1), one replaces the
objective function with the definition of the Division 1 efficiency, and additionally
requires the system efficiency to equal E0 in the constraint set, to get the following
model:
hð1Þ ð1Þ
ð1Þ g = 1 wg Z g0
E0 = max: m
i = 1 vi X i0
s m
s:t: ur Y r0 = E 0 vi X i0
r=1 i=1
hð1Þ m
ð1Þ
wg Z gj - vi X ij ≤ 0, j = 1, . . . , n
g=1 i=1
hðkÞ hðk - 1Þ
ðkÞ ðk - 1Þ
wg Z gj - wg Z gj ≤ 0, k = 2, . . . , p - 1, j = 1, . . . , n
ðk - 1Þ ðk - 2Þ
g=h þ1 g=h þ1
s hðp - 1Þ
ðp - 1Þ
ur Y rj - wg Z gj ≤ 0, j = 1, . . . , n
r=1 g=h ðp - 2Þ
þ1
ur , vi , wg ≥ ε, 8r, i, g
ð12:3Þ
The efficiencies of Division 1 measured from this model for all DMUs are then
comparable.
At this stage, if the efficiencies of Division 2 of all DMUs are to be compared,
then Model (12.3) is modified by replacing the objective function with the definition
of the Division 2 efficiency, and adding the constraint of requiring the Division
ð1Þ
1 efficiency to equal E 0 to obtain the following model:
h ð 2Þ ð2Þ
ð2Þ g = hð1Þ þ1 wg Z g0
E0 = max:
hð1Þ ð1Þ
g = 1 wg Z g0
s m
s:t: ur Y r0 = E0 vi X i0
r=1 i=1
hð1Þ m
ð1Þ ð1Þ
wg Z g0 = E 0 vi X i0
g=1 i=1
hð1Þ m
ð1Þ
wg Z gj - vi X ij ≤ 0, j = 1, . . . , n
g=1 i=1
hðkÞ hðk - 1Þ
ðkÞ ðk - 1Þ
wg Z gj - wg Z gj ≤ 0, k = 2, . . . , p - 1, j = 1, . . . , n
ðk - 1Þ ðk - 2Þ
g=h þ1 g=h þ1
s hðp - 1Þ
ðp - 1Þ
ur Y rj - wg Z gj ≤ 0, j = 1, . . . , n
r=1 g = hðp - 2Þ þ1
ur , vi , wg ≥ ε, 8r, i, g
ð12:4Þ
This process can be continued until all divisions have been enumerated. The
efficiencies of each division among different DMUs are then comparable. One
technical problem to be noted in solving Model (12.4) is that requiring the two
equality constraints to be satisfied may cause infeasibility due to rounding errors.
The second equality constraint may need to be split into two inequality ones to be
ð1Þ
feasible, one with a slightly larger value of E0 as the upper bound and the other with
ð1Þ
a slightly smaller value of E 0 as the lower bound.
In contrast to efficiency decomposition, where the system efficiency is expressed as

the ratio of the aggregation of its outputs to that of its inputs, efficiency aggregation
defines the system efficiency as an aggregation of the division efficiencies according
to a pre-specified mathematical relation. The additive form that defines the system
efficiency as a weighted average of the division efficiencies is usually adopted; that
ðk Þ ðk Þ
is, E0 = pk = 1 ωðkÞ E 0 , where E0 are defined in Eq. (12.2). The weights ω(k) can be
either constants, pre-specified by the decision maker to represent the importance of
each division, or variables, to be reflected from the data. If it is the former, then the
model is:
hð1Þ ð1Þ p-1 hðkÞ ðk Þ

ð1Þ g = 1 wg Z g0 ðk Þ g = hðk - 1Þ þ1 wg Z g0
max:ω m þ ω þ
hðk - 1Þ ðk - 1Þ
i = 1 vi X i0 k=2 g = hðk - 2Þ þ1 wg Z g0
s ð12:5Þ
r = 1 ur Y r0
ωðpÞ ðp - 1Þ
h ðp - 1 Þ
g = hðp - 2Þ þ1 wg Z g0
s:t: Constraints in Model ð12:1Þ
In this case the objective function is nonlinear.

If it is the latter that the weights are to be reflected from the data, then a suitable
selection of the function form to express the weights is able to convert the model to a
linear one. Similar to the basic two-stage case, if we define the weights to be the
proportion of the aggregate input consumed by a division in that consumed by all
divisions, i.e.,
m
i = 1 vi X i0
ωð1Þ = m h
i = 1 i i0 þ
v X g = 1 wg Z g0
hðk - 1Þ ðk - 1Þ ð12:6Þ
ðk Þ g = hðk - 2Þ þ1 wg Z g0
ω = m h
k = 2, . . . , p
i = 1 vi X i0 þ g = 1 wg Z g0
p ðk Þ ðk Þ
then the weighted average becomes k = 1 ω E0 =
s h m h p ðk Þ
r = 1 ur Y r0 þ g = 1 wg Z g0 = i = 1 vi X i0 þ g = 1 wg Z g0 , with k = 1ω =1
and ω ≥ 0, k = 1, . . ., p. Note here the
(k)
superscript (k) in Z gðkÞ
is discarded because it
is not necessary. The associated model for measuring the efficiencies is then:
s h
r = 1 ur Y r0 þ g = 1 wg Z g0
max: m
i = 1 vi X i0 þ
h
g = 1 wg Z g0
ð12:7Þ
s:t: Constraints in Model ð12:1Þ
variable substitution technique presented in Charnes and Cooper (1962). Solving
this model obtains an optimal solution (u*, v*, w*) that yields the highest system
efficiency. If there are multiple solutions, then the procedure for determining a set of
division efficiencies discussed in the preceding subsection can be applied. Based on
Eq. (12.6), the optimal weight associated with each division is calculated.
Model (12.7) is applied to the same example in Table 12.1 in order to calculate the
system and division efficiencies, with the results shown in Table 12.2. There are
several points to be noted. First, the system efficiency for every DMU contained in
column two is higher than that measured from the efficiency decomposition model,
contained in column seven of Table 12.1. The reason for this is similar to that
discussed in Chap. 11, and this will be discussed in more detail in Sect. 12.3. Second,
the efficiencies of the four divisions of every DMU are the same as those measured
from the efficiency decomposition model and contained in Table 12.1. This is
because there is only one intermediate product for every division, and the solution
in this case is unique. The numbers in parentheses are the weights associated with the
corresponding division calculated from Eq. (12.6). It is easy to verify that the system
efficiency of each DMU is indeed the weighted average of the division efficiencies.
For example, the weighted average of the division efficiencies of DMU A is
(4/9) × (1/2) + (2/9) × (1) + (2/9) × (1/2) + (1/9)× (1/2) = 11/18, which is just the
efficiency of the system in column two. Different from expected, the division with
the highest efficiency does not necessarily have the largest weight. The reason is
perhaps that we need the aggregate effect of the weight and the efficiency for all four
Table 12.2 Efficiencies and weights calculated from the efficiency aggregation model for the basic
series example
DMU E0 E0 ω(1) E0 ω(2) E0 ω(3) E0 ω(4)
A 11/18 1/2 (4/9) 1 (2/9) 1/2 (2/9) 1/2 (1/9)
B 3/4 1 (1/3) 2/3 (1/3) 1/2 (2/9) 3/4 (1/9)
C 11/28 1/3 (9/14) 1/3 (3/14) 1 (1/14) 1/2 (1/14)
D 17/28 3/4 (12/28) 4/9 (9/28) 3/4 (4/28) 1/3 (3/28)
E 7/12 1/2 (1/2) 2/3 (1/4) 1/2 (1/6) 1 (1/12)
divisions to have the highest system efficiency, and a larger weight for the divi-
sion may have the associated multipliers that do not produce the highest system
efficiency for the system when the efficiencies of all four divisions are aggregated.
12.2 Independent Efficiency Measures
The basic series system is important from a theoretical point of view. In reality,
however, it is rare to find systems that do not need other inputs to process the
intermediate products to become the final outputs. Besides, there are also cases in
which some of the intermediate products are sold as spare parts. A more realistic
structure is to allow each division to consume exogenous inputs and produce
exogenous outputs to be a general series system.
Figure 12.3 depicts the structure of the general series system, where p divisions
ðk Þ
are connected in series. Each division consumes the exogenous inputs X i supplied
ðk - 1Þ
from outside and endogenous inputs Z g produced by the preceding division to
produce the exogenous outputs Y ðrkÞ to send to outside, and endogenous outputs Z ðgkÞ
for the succeeding division to use. The first and last divisions are a little different, in
that the former consumes only the exogenous inputs and the latter produces only the
exogenous outputs. Similar to the case of the intermediate product Z gðkÞ , the sub-
scripts i and r for the input and output of every division run the same from 1 to m and
1 to s, respectively, for simplicity of notation. As a matter of fact, they are usually
different for each division in real-world applications.
Several models have been proposed to measure the efficiency of this type of
system. The simplest one is the independent model that treats each division as an
independent DMU and measures their efficiencies separately. In other words, p DEA
X i( k ) , i = m( k –1) + 1, ..., m( k ) X i( p ) , i = m( p –1) + 1, ..., m( p )
X i(1) Z g(1) Z g( k – 1) Z g(k ) Z g( p –1) Yr( p )

1 … k … p
i = 1, ..., m(1) g = 1, ..., h(1) g = h ( k –2) +1, ..., h ( k –1)
( p –2 )
g = h( k –1) +1, ..., h( k ) g = h +1, ..., h
( p –1)
r = s ( p–1) +1, ..., s ( p )
Yr(1) , r = 1, ..., s (1) Yr( k ) , r = s ( k –1) +1, ..., s ( k )
Fig. 12.3 Structure of the general series system

12.2 Independent Efficiency Measures 299
models are required to measure the efficiency of the p divisions. Under constant
returns to scale, the p models are:
Division 1
sð1Þ ð1Þ h ð 1Þ ð1Þ

ð1Þ r = 1 ur Y r0 þ g = 1 wg Z g0
E0 = max: m ð 1Þ ð1Þ
i = 1 vi X i0
sð1Þ hð1Þ mð1Þ ð12:8Þ
ð1Þ ð1Þ ð1Þ
r=1 g=1 i=1
ur , vi , wg ≥ ε, 8r, i, g
Division k, k = 2,. . ., p-1
sðkÞ ðk Þ hðkÞ ðk Þ
ðk Þ r = sðk - 1Þ þ1 ur Y r0 þ g = hðk - 1Þ þ1 wg Z g0
E0 = max:
m ðk Þ ðk Þ h ð k - 1Þ ðk - 1Þ
i = mðk - 1Þ þ1 vi X i0 þ g = hðk - 2Þ þ1 wg Z g0
sðkÞ hðkÞ
ðkÞ ðkÞ
s:t: ur Y rj þ wg Z gj -
r = sðk - 1Þ þ1 g = hðk - 1Þ þ1
mðkÞ
hðk - 1Þ
ðkÞ ðk - 1Þ
vi X ij þ wg Z gj ≤ 0, j = 1, . . . , n
i = mðk - 1Þ þ1 g=h ðk - 2Þ
þ1
ur , vi , wg ≥ ε, 8r, i, g
Division p
sðpÞ ðpÞ
ðpÞ r = sðp - 1Þ þ1 ur Y r0
E0 = max: ð - Þ
m ð pÞ ðp Þ h p 1 ðp - 1Þ
i = mðp - 1Þ þ1 vi X i0 þ g = hðp - 2Þ þ1 wg Z g0
sðpÞ mðpÞ hðp - 1Þ

ðpÞ ðpÞ ðp - 1Þ
s:t: ur Y rj - vi X ij þ wg Z gj ≤ 0, j = 1, . . . , n
r = sðp - 1Þ þ1 i = mðp - 1Þ þ1 g = hðp - 2Þ þ1
ur , vi , wg ≥ ε, 8r, i, g
The results show the efficiency of each division measured as an independent

DMU. This set of models is the basis for other models to measure efficiencies, and
the only weakness is its inability to show the efficiency of the system.
The ratio-form efficiency measurement models are relatively easier to understand

and formulate. Similar to the discussion in Sect. 12.1, there are two ideas used in
formulating the ratio-form models to measure the efficiency of systems with the
general series structure: efficiency aggregation and decomposition.
Based on the structure of Fig. 12.3, the efficiencies of the p divisions under constant
returns to scale are:
sð1Þ ð1Þ h ð 1Þ ð1Þ

ð1Þ r = 1 ur Y r0 þ g = 1 wg Z g0
E0 = m ð 1Þ ð1Þ
i = 1 vi X i0
sðkÞ ðk Þ hðkÞ ðk Þ
ðk Þ r = sðk - 1Þ þ1 ur Y r0 þ g = hðk - 1Þ þ1 wg Z g0
E0 = k = 2, . . . , p - 1 ð12:9Þ
mðkÞ ðk Þ hðk - 1Þ ðk - 1Þ
i = mðk - 1Þ þ1 vi X i0 þ g = hðk - 2Þ þ1 wg Z g0
sðpÞ ðpÞ
ðpÞ r = sðp - 1Þ þ1 ur Y r0
E0 =
mðpÞ ðpÞ hðp - 1Þ ðp - 1Þ
i = mðp - 1Þ þ1 vi X i0 þ g = hðp - 2Þ þ1 wg Z g0
The efficiency aggregation model, in additive form, defines the system efficiency
as a weighted average of the division efficiencies. To make the model linear, the
weight associated with each division is defined as the proportion of the aggregate
input, including exogenous and endogenous inputs, consumed by this division in
that consumed by all divisions. The total aggregate input of all divisions is
m h
i = 1 vi X i0 þ g = 1 wg Z g0 , and the weights associated with the divisions are:
m ð 1Þ ð1Þ
i=1 vi X i0
ωð1Þ = m h
i=1 vi X i0 þ g=1 wg Z g0
ð12:10Þ
mðkÞ ðk Þ hðk -1Þ ðk -1Þ
i=mðk -1Þ þ1 vi X i0 þ g=hðk -2Þ þ1 wg Z g0
ωðkÞ = m h
k =2, ...,p-1
i=1 vi X i0 þ g=1 wg Z g0
with pk = 1 ωðkÞ = 1 and ω(k) ≥ 0, k = 1, . . ., p. Based on Eqs. (12.9) and (12.10), the
p ðk Þ ðk Þ
weighted average of the p division efficiencies is k = 1 ω E0 =
s h m h
r = 1 ur Y r0 þ g = 1 wg Z g0 = i = 1 vi X i0 þ g = 1 wg Z g0 . The associated effi-
ciency aggregation model under constant returns to scale is then:
s h
r = 1 ur Y r0 þ g = 1 wg Z g0
max: m h
i = 1 vi X i0 þ g = 1 wg Z g0
sð1Þ hð1Þ mð1Þ
ð1Þ ð1Þ ð1Þ
r=1 g=1 i=1
sðkÞ hðkÞ
ðkÞ ðkÞ
ur Y rj þ wg Z gj -
r = sðk - 1Þ þ1 g=h ðk - 1Þ
þ1
mðkÞ hðk - 1Þ
ðkÞ ðk - 1Þ
vi X ij þ wg Z gj ≤ 0, k = 2, . . . , p - 1, j = 1, . . . , n
i = mðk - 1Þ þ1 g = hðk - 2Þ þ1
sðpÞ mðpÞ hðp - 1Þ

ðpÞ ðpÞ ðp - 1Þ
r = sðp - 1Þ þ1 i = mðp - 1Þ þ1 g=h ðp - 2Þ
þ1
ur , vi , wg ≥ ε, 8r, i, g
ð12:11Þ
At optimality, the system efficiency is the objective value, and the division
efficiencies are calculated according to Eq. (12.9), with the weights calculated
from Eq. (12.10). Divisions with lower efficiencies are the ones that have greater
effects on the performance of the system, whose operations should be carefully
monitored in order to achieve better results.
The efficiency decomposition model uses the inputs and outputs of the system to
measure the system efficiency, taking the operations of the divisions into account.
The model is:
s
r = 1 ur Y r0
E 0 = max: m
i = 1 vi X i0 ð12:12Þ
The most important task in efficiency decomposition is to find the relationship

between the system and division efficiencies, and there are at least two ways to
derive this for systems with the general series structure.
The easier one is based on the weighted average of the division efficiencies,
which can be derived as:
s h
r = 1 ur Y r0 þ
p
ðk Þ g = 1 wg Z g0 E0 þ ρ
ωðkÞ E 0 = = ð12:13Þ
m h 1þρ
k=1 i = 1 vi X i0 þ g = 1 wg Z g0
h m
where ρ = g = 1 wg Z g0 = i = 1 vi X i0 . Rearranging the terms obtains:
p p p
E0 = ð1 þ ρÞ ωðkÞ E 0 - ρ = ωðkÞ E 0 - ρ 1 - ωðkÞ E 0
k=1 k=1 k=1
p
ðk Þ
≤ ωðkÞ E 0 ð12:14Þ
k=1
This equation indicates that the system efficiency is a weighted average of the
division efficiencies adjusted by a factor ρ. Further, the system efficiency E0 will
ðk Þ
equal the weighted average of the division efficiencies pk = 1 ωðkÞ E0 in two cases,
ðk Þ
either E0 and all E0 are equal to one or ρ is equal to zero. In the former case, when
all divisions are efficient, the system will obviously be efficient. Mathematically,
ðk Þ
when all divisions are efficient, we have pk = 1 ωðkÞ E 0 = 1, which leads to E 0 =
p ðk Þ ðk Þ
ð1 þ ρÞ × k = 1 ω E0 - ρ = 1: For the latter case, ρ = 0 indicates that
h
= 0, which implies that every division is operating independently,
g = 1 wg Z g0
without linkages connecting any two divisions. The system efficiency is obviously
the weighted average of the independent division efficiencies. Another property to
ðk Þ
be noted is that since ρ ≥ 0, according to Eq. (12.13), we have pk = 1 ωðkÞ E 0 =
ðE 0 þ ρÞ=ð1 þ ρÞ ≥ E0 : This indicates that the system efficiency is always less than
or equal to the weighted average of the division efficiencies.
Another way to decompose the system efficiency, based on the idea presented in
Kao (2014a), is to transform the general structure into a basic series structure of
p subsystems, where each subsystem is composed of one original and one dummy
division connected in parallel. With this transformation, all exogenous inputs and
outputs are carried by the dummy divisions to become endogenous. Figure 12.4
shows this transformation, where the real divisions are represented by squares and
the dummy ones by circles. Dummy divisions let the inputs enter and leave without
ð2Þ ðpÞ
doing anything. For example, X i , . . . , X i enter dummy Division 1, and the same
ð2Þ ðpÞ
factors X i , . . . , X i come out of this dummy division. Since the inputs and outputs
are the same, the efficiencies of dummy divisions are equal to one, and the con-
straints corresponding to dummy divisions are identities and can be deleted. Using
Subsystem II as an example to explain this, the constraint associated with the real
s ð 2Þ ð2Þ hð2Þ ð2Þ
Division 2 is the same as before: r = sð1Þ þ1 ur Y rj þ g = hð1Þ þ1 wg Z gj -
m ð 2Þ ð2Þ hð1Þ ð1Þ
i = mð1Þ þ1 vi X ij þ g = 1 wg Z gj ≤ 0, and that associated with the dummy Divi-
s ð 1Þ ð1Þ p m ðk Þ ðk Þ
sion 2 is r = 1 ur Y rj þ k=3 i = mðk - 1Þ þ1 vi X ij -
sð1Þ ð1Þ p m ðk Þ ðk Þ
r = 1 ur Y rj þ k=3 i = mðk - 1Þ þ1 vi X ij ≤ 0: When all p subsystems are consid-
ered, the constraints corresponding to this transformed system are the same as the
original system, Model (12.12). The efficiency of each subsystem is its aggregate
output divided by its aggregate input, which, referring to Fig. 12.4, is:
I II
X i(1) Z g(1) X i( 2 ) Z g( 2)
1 2
Yr(1) Z (1)
g Yr( 2 )
X i(1) , ..., X i( p )
…
( p)
X i( 2) , ..., X i( p ) X i( 2) , ..., X i( p ) X i
( 3)
, ..., X i X i(3) , ..., X i( p )
1 2
Y r
(1)
Yr(1)
P- 1 P
X i( p–1) Z g( p–1) X i( p )
( p– 2)
p –1 p
Z g Yr( p –1) Z g( p–1) Yr( p )
… Yr(1) , ..., Yr( p )
X i( p ) X i( p) p
p–1
Yr(1) , ..., Yr( p– 2) Yr(1) , ..., Yr( p–2) Yr(1) , ..., Yr( p –1) Yr(1) , ..., Yr( p –1)
Fig. 12.4 Series-parallel transformation of the general series system
sð1Þ hð1Þ p mðkÞ

ð1Þ ð1 Þ ðk Þ
ur Y r0 þ wg Z g0 þ vi X i0
r=1 g=1 k = 2 i = mðk - 1Þ þ1
E I0 = m
vi X i0
i=1
k sð d Þ hðkÞ p mðdÞ
ðd Þ ðk Þ ðd Þ
d = 1 r = sðd - 1Þ þ1 g=h ð k - 1Þ
þ1 d = kþ1 i = mðd - 1Þ þ1
E K0 = , K = II, . . . , P - 1
p mðdÞ hðk - 1Þ k-1 sðdÞ
ðd Þ ðk - 1Þ ðd Þ
vi X i0 þ wg Z g0 þ ur Y r0
d = k i = mðd - 1Þ þ1 g = hðk - 2Þ þ1 d = 1 r = sðd - 1Þ þ1
s
ur Y r0
r=1
E P0 =
mðpÞ hðp - 1Þ p-1 sð k Þ
ðp Þ ðp - 1 Þ ðk Þ
i = mðp - 1Þ þ1 g=h ðp - 2Þ
þ1 k = 1 r = sðk - 1Þ þ1
It is noted that the numerator of the efficiency of a subsystem is just the

denominator of that of its succeeding subsystem. The product of all subsystem
efficiencies thus becomes the system efficiency:
P K s m
K=I 0E = u Y
r = 1 r r0 = v X
i = 1 i i0 :.
Since each subsystem is composed of two independent divisions, one real and one
dummy, its efficiency is the weighted average of these two divisions, where the
weight is the proportion of the aggregate input consumed by the division in that
consumed by the two divisions. Using Subsystem II to explain this, the total
m ð 2Þ ð2Þ hð1Þ ð1Þ
aggregate input of this subsystem is i = mð1Þ þ1 vi X i0 þ g = 1 wg Z g0 þ
p mðkÞ ðk Þ sð1Þ ð1Þ p m ðk Þ ðk Þ
k=3 i = mðk - 1Þ þ1 vi X i0 þ r = 1 ur Y r0 , or k=2 i = mðk - 1Þ þ1 vi X i0 þ
ð 1Þ ð1Þ ð 1 Þ ð1Þ
h s
g = 1 wg Z g0 þ r = 1 ur Y r0 , and the aggregate input consumed by the
m ð 2Þ ð2Þ h ð 1Þ ð1Þ
real division 2 is i = mð1Þ þ1 vi X i0 þ g = 1 wg Z g0 , the weight associated with
mð2Þ ð2Þ h ð 1Þ ð1Þ
the real division 2 is thus ωð2Þ = i = mð1Þ þ1 vi X i0 þ g = 1 wg Z g0 =
p m ðk Þ ðk Þ h ð 1Þ ð1Þ sð1Þ ð1Þ
k=2 i = mðk - 1Þ þ1 vi X i0 þ g = 1 wg Z g0 þ r = 1 ur Y r0 , and that associated
with the dummy division 2 is 1 - ω . The weighted average of the two divisions is:
(2)
sð2Þ hð2Þ sð1Þ p mðkÞ

ð2Þ ð2Þ ð1Þ ðkÞ
ur Y r0 þ wg Z g0 ur Y r0 þ vi X i0
r =sð1Þ þ1 ð1Þ
g = h þ1 r=1 k=3 i= mðk - 1Þ þ1
ωð2Þ × þ ð1- ωð2Þ Þ×
mð2Þ hð1Þ sð1Þ p mðkÞ
ð2Þ ð1Þ ð1Þ ðkÞ
vi X i0 þ wg Z g0 ur Y r0 þ vi X i0
i =mð1Þ þ1 g=1 r=1 k = 3 i= mðk - 1Þ þ1
2 sðkÞ hð2Þ p mðkÞ

ðkÞ ð2Þ ðkÞ
k = 1 r = sðk - 1Þ þ1 g = h þ1 ð1Þ k = 3 i = mðk - 1Þ þ1
= = E II0
p mðkÞ hð1Þ sð1Þ
ðkÞ ð2Þ ð1Þ
k =2 i = mðk - 1Þ þ1 g=1 r=1
which is just the efficiency of Subsystem II. This relationship holds for all sub-
systems. The system efficiency can then be decomposed as:
P p p
ðk Þ ðk Þ
E0 = EK0 = ω ð k Þ E 0 þ 1 - ωð k Þ ≥ E0 ð12:15Þ
K=I k=1 k=1
ðk Þ ðk Þ
The last inequality is due to E 0 ≤ 1, which leads to ωðkÞ E0 þ 1 - ωðkÞ ≥
ðk Þ ðk Þ ðk Þ ðk Þ
ω E 0 þ 1 - ωðkÞ × E 0 = E0 : The larger the weight ω(k), the closer the system
is to a basic series structure, and the closer the system efficiency is to the product of
the division efficiencies. Combining the relationships in Eqs. (12.14) and (12.15),
the following relationship is obtained:
p p
ðk Þ ðk Þ
E0 ≤ E0 ≤ ωð k Þ E 0
k=1 k=1
In other words, the system efficiency lies between the product and weighted
average of the division efficiencies.
Recall the example of the basic series structure discussed in Sect. 12.1, where the
system efficiency is E0 = uY0/vX0, with the results of the five DMUs shown in
column seven of Table 12.1. The weighted average of the division efficiencies is
4 ðk Þ ðk Þ 3 3
k = 1 ω E0 = uY 0 þ g = 1 wg Z g0 = vX 0 þ g = 1 wg Z g0 , with the results
shown in column two of Table 12.2. Since the weighted average of the division
4 ðk Þ ðk Þ
efficiencies can be expressed as k = 1 ω E 0 = ðE 0 þ ρÞ=ð1 þ ρÞ ≥ E 0 , where
ρ = 3g = 1 wg Z g0 =vX 0 , the weighted average is strictly larger than E0 for ρ > 0.
The values in column seven of Table 12.1 and those in column two of Table 12.2
verify this relationship. Finally, no matter whether the efficiency decomposition or
aggregation model is used, the divisions that have lower efficiencies are the major
ones that cause the inefficiency of the system.
The ratio-form models have the merit of being able to calculate the division
efficiencies when measuring the efficiency of the system. However, they are not
able to show the targets of the factors for inefficient DMUs to become efficient. If the
targets are desired, then the distance function models must be used. The distance
parameters associated with the system and divisions have been proposed in the
literature, and these will be discussed in this section.
The distance function can be defined from either the input or output side, and here we
will only discuss the input-oriented model, as the output-oriented model can be
explained in a similar manner.
Based on the structure of the general series system shown in Fig. 12.3, the
distance function model under constant returns to scale, with the distance parameter
attached to only the exogenous inputs, can be formulated as:
min:θ
n
ðkÞ ðkÞ ðkÞ
s:t: λj X ij ≤ θX i0 , i = mðk - 1Þ þ 1, . . . , mðkÞ , k = 1, . . . , p
j=1
n
ðkÞ ðkÞ ðkÞ
λj Z gj ≥ Z g0 , Z g = hðk - 1Þ þ 1, . . . , hðkÞ , k = 1, . . . , p - 1
j=1
n
ðkþ1Þ ðkÞ ðkÞ
λj Z gj ≤ Z g0 , g = hðk - 1Þ þ 1, . . . , hðkÞ , k = 1, . . . , p - 1
j=1
n
ðkÞ ðkÞ ðkÞ
λj Y rj ≥ Y r0 , r = sðk - 1Þ þ 1, . . . , sðkÞ , k = 1, . . . , p
j=1
ðkÞ
λj ≥ 0, j = 1, . . . , n, k = 1, . . . , p
ð12:16Þ
Note that the non-Archimedean number ε has been omitted for simplicity of
ðk Þ
expression. The convexity constraint of nj = 1 λj = 1 should be added if division
k is assumed to have a variable returns to scale technology.
Note that Model (12.16) assumes that each division consumes different inputs
and produces different outputs. If Xf is consumed by two or more divisions, for
example Divisions a and b, then the constraints for Divisions a and b must be
n ðaÞ ðaÞ n ðbÞ ðbÞ ðaÞ ðbÞ
combined to become j = 1 λj X fj þ j = 1 λj X fj ≥ X f 0 þ X f 0 : This way of
formulation also applies to the case of common outputs produced by two or more
divisions. At optimality, the targets for inputs Xi, intermediate products Zg, and
outputs Yr, based on the constraints of Model (12.16), are:
ðk Þ n
ðk Þ ðk Þ
X i0 = λj X ij , i = mðk - 1Þ þ 1, . . . , mðkÞ , k = 1, . . . , p
j=1
ðkÞðoutÞ n ðkÞðinÞ n
ðkÞ ðk Þ ðkþ1Þ ðk Þ
Z g0 = λj Z gj , Z g0 = λj Z gj , g = 1, . . . , h, k = 1, . . . , p - 1
j=1 j=1
ðk Þ n
ðk Þ ðk Þ
Y r0 = λj Y rj , r = sðk - 1Þ þ 1, . . . , sðkÞ , k = 1, . . . , p
j=1
ð12:17Þ
ðkÞðoutÞ
Note that the target value of the intermediate product Z g0 , considered as the
ðk ÞðinÞ
output of Division k, may not be the same as considered as the input of Z g0 ,
Division k + 1. However, the constraints of Model (12.16) ensure that the amount
ðk ÞðoutÞ
being produced by Division k (i.e., Z g0 ) will be greater than or equal to the
ðk Þ
observed amount of Z g0 , which in turn must be greater than or equal to the amount to
ðk ÞðinÞ
be used by Division k + 1 (i.e., Z g0 ).
As a comparison to this model, consider the efficiency decomposition model
(12.12). The dual of the linearized Model (12.12) can be formulated as:
min:θ
n
ðkÞ ðkÞ ðkÞ
j=1
n n
ðkÞ ðkÞ ðkþ1Þ ðkÞ
λj Z gj ≥ λj Z gj , g = hðk - 1Þ þ 1, . . . , hðkÞ , k = 1, . . . , p - 1
j=1 j=1
n
ðkÞ ðkÞ ðkÞ
j=1
ðkÞ
λj ≥ 0, j = 1, . . . , n, k = 1, . . . , p
ð12:18Þ
This model is similar to Model (12.16), except the two constraints of

n ðk Þ ðkÞ
ðk Þ ðkþ1Þ ðk Þ ðk Þ
≥ Z g0 and nj = 1 λj
j = 1 λj Z gj Z gj ≤ Z g0 in Model (12.16) have been com-
ðk Þ ðk Þ ðkþ1Þ ðk Þ
bined into one constraint of nj = 1 λj Z gj ≥ nj = 1 λj Z gj in Model (12.18). In
this case Model (12.18) is less stringent than Model (12.16), due to the aggregation
of two constraints to one. The system efficiency measured from Model (12.18) will
be less than or equal to that measured from Model (12.16). This relationship is also
clear from the dual side, in that the dual of Model (12.16) allows the multipliers
associated with the intermediate products considered as the outputs of a division to
be different from those associated with the same intermediate products yet consid-
ered as the inputs of the succeeding division, while the dual of Model (12.18)
requires them to be the same. The dual solution of Model (12.16) is thus greater
than or equal to that of Model (12.18).
The distance function models are able to show the targets of the factors that should
be met to enable inefficient DMUs to become efficient. However, they are not able to
measure the efficiency of the component divisions. To address this weakness an
intuitive idea is to assign different distance parameters to each division, with the
expectation that these can represent the efficiencies of the corresponding divisions.
Let θ(k) be the distance parameter corresponding to the kth division, and the
weighted average of the division parameters pk = 1 ωðkÞ θðkÞ be the efficiency of the
system, where the weights ω(k) are pre-specified constants. The model, from the input
side and under constant returns to scale, is as follows:
p
min: ωðkÞ θðkÞ
k=1
n
ðkÞ ðkÞ ðkÞ
s:t: λj X ij ≤ θðkÞ X i0 , i = mðk - 1Þ þ 1, . . . , mðkÞ , k = 1, . . . , p
j=1
n n
λj Z gj ≥ λj Z gj , g = hðk - 1Þ þ 1, . . . , hðkÞ , k = 1, . . . , p - 1
j=1 j=1
n
ðkÞ ðkÞ ðkÞ
j=1
θðkÞ ≤ 1, k = 1, . . . , p
ðkÞ
λj ≥ 0, j = 1, . . . , n, k = 1, . . . , p
ð12:19Þ
The constraints are essentially the same as those of Model (12.18), except that
different divisions are allowed to have different (input) distance parameters. If all θ(k)
are required to be the same, then this model boils down to Model (12.18). This model
is thus more flexible, and the system efficiency calculated from it will be less than or
equal to that calculated from Model (12.18).
To get some idea of the meaning of θ(k), we further allow each factor to have a
ðk Þ ðk Þ ðk Þ ðk Þ - ðk Þ
different parameter θi and define θi = X i0 - si =X i0 : The constraint can
then be expressed as:
n
ðk Þ ðk Þ ðk Þ ðk Þ ðk Þ ðk Þ -
λj X ij = θi X i0 = X i0 - si , i = mðk - 1Þ þ 1, . . . , mðkÞ , k = 1, . . . , p
j=1
ðk Þ ðk Þ
The parameter θi becomes the efficiency of the factor Xi. If the θi
corresponding to the factors of the same division are required to be the same, denoted
as θ(k), then θ(k) may represent an efficiency measure of this division.
To further investigate the meaning of θ(k), consider the dual of this model, which
can be formulated as:
p sðkÞ p
ðkÞ
max: ur Y r0 - αðkÞ
k =1 r=sðk -1Þ þ1 k =1
mðkÞ
ðkÞ
s:t: vi X i0 -αðkÞ =ωðkÞ , k=1, ...,p
i=mðk -1Þ þ1
ð1Þ ð1Þ ð1Þ
ur Y r j þ wg Z g j - vi X i j ≤0, j=1, ...,n
r=1 g=1 i=1
sðkÞ hðkÞ mðkÞ hðk -1Þ
ðkÞ ðkÞ ðkÞ ðk-1Þ
ur Y r j þ wg Z g j - vi X i j þ wg Z g j ≤0, k=2,...,p-1 j=1, ...,n
r=sðk -1Þ þ1 g=h ðk -1Þ
þ1 i=mðk -1Þ þ1 g=h ðk -2Þ
þ1
sðpÞ mðpÞ hðp-1Þ
ðpÞ ðpÞ ðp-1Þ
ur Y r j - vi X i j þ wg Z g j ≤0, j=1, ...,n
r=sðp-1Þ þ1 i=mðp-1Þ þ1 g=h ðp-2Þ
þ1
ur , vi , wg , αðkÞ ≥ε, 8r, i, g, k
ð12:20Þ
Substituting the expression of α(k) in the first constraint set into the objective
p sðkÞ ðk Þ p m ðk Þ ðk Þ
function obtains k=1 r = sðk - 1Þ þ1 ur Y r0 - k=1 i = mðk - 1Þ þ1 vi X i0 þ 1, which
indicates that Model (12.19) is an additive model, rather than a radial one. We will
use an example to show that the division efficiencies calculated from this model,
which has the form of Eq. (12.9), are not the same as θ(k)measured from Model
(12.19).
While the notations in Models (12.19) and (12.20) are quite complicated to
understand, in real applications this is rather simple. Consider a simple example of
three divisions, as depicted in Fig. 12.5, where X1 is used by Division 1 to produce
the exogenous output Y1 and intermediate product Z1. This intermediate product,
together with the exogenous input X2, is then used by Division 2 to produce the
Fig. 12.5 An example X2 X3

system with three divisions
Z1 Z2
X1 1 2 3 Y3
Y1 Y2
exogenous output Y2 and another intermediate product Z2. Finally, Division 3 applies
the intermediate product Z2 and the exogenous input X3 to produce the exogenous
output Y3. The input, output, and intermediate product of every division are different.
Models (12.19) and (12.20) for this example become:
min:ωð1Þ θð1Þ þ ωð2Þ θð2Þ þ ωð3Þ θð3Þ

n
ð1Þ
s:t: λj X 1j ≤ θð1Þ X 10
j=1
n
ð2Þ
j=1
n
ð3Þ
j=1
n n
ð1Þ ð2Þ
j=1 j=1
n n
ð2Þ
λj Z 2j ≥
ð3Þ
λj Z 2j ð12:21Þ
j=1 j=1
n
ð1Þ
λj Y 1j ≥ Y 10
j=1
n
ð2Þ
λj Y 2j ≥ Y 20
j=1
n
ð3Þ
λj Y 3j ≥ Y 30
j=1
θðkÞ ≤ 1, k = 1, 2, 3
ðkÞ
λj ≥ 0, k = 1, 2, 3, j = 1, . . . , n
and
max:u1 Y 10 þ u2 Y 20 þ u3 Y 30 - αð1Þ - αð2Þ - αð3Þ

s:t: v1 X 10 - αð1Þ = ωð1Þ
v2 X 20 - αð2Þ = ωð2Þ
v3 X 30 - αð3Þ = ωð3Þ
u1 Y 1j þ w1 Z 1j - v1 X 1j ≤ 0, j = 1, . . . , n
u2 Y 2j þ w2 Z 2j - ðv2 X 2j þ w1 Z 1j Þ ≤ 0, j = 1, . . . , n
u3 Y 3j - ðv3 X 3j þ w2 Z 2j Þ ≤ 0, j = 1, . . . , n
ðkÞ
ur , vi , wg , α ≥ 0, r = 1, 2, 3, i = 1, 2, 3, g = 1, 2, k = 1, 2, 3
ð12:22Þ
which are quite straightforward.

Columns two to nine of Table 12.3 show the data for five DMUs with the
structure shown in Fig. 12.5. The division efficiencies calculated from Model
(12.21) are shown in columns ten to twelve, and the objective value, which is the
average of the three-division efficiencies, is shown in column thirteen under the
heading E0. The division efficiencies calculated from Model (12.22) are shown in the
last three columns. Note that Models (12.21) and (12.22) are the primal and dual of
each other, and their objective values are the same. The objective value of Model
(12.22), which is the system efficiency, is thus the same as that calculated from
Model (12.21), shown in column thirteen under the heading E0. It is also noted that
the division efficiencies calculated from these two models are not exactly the same.
While the system efficiency is the average of the division efficiencies in Model
(12.21), it is not the case for Model (12.22). θ(k) is thus another way of measuring
The distance function models are either input- or output-oriented, which fix the
level of either outputs or inputs to find the best level of the other. If the excessively
used inputs are desired to be reduced and the insufficiently produced outputs are to
be expanded at the same time, then the directional distance function approach can be
applied. Model (11.25), developed for general two-stage systems, can then be
extended to general multi-stage systems.
It is usually desirable to reduce the excessive inputs that have been used while at the
same time increasing the insufficient outputs that should have been produced, and
the slacks-based approach is able to achieve this goal.
The idea of the slacks-based approach is to use the average distance parameter of
all input factors divided by that of all output factors to be the system efficiency,
where the slack variables are the difference between the targets and the observations,
expressed as:
Table 12.3 Data and efficiencies for the three-division example

DMU X1 Y1 Z1 X2 Y2 Z2 X3 Y3 θ1 θ2 θ3 E0 E10 E20 E 30
A 2 1 2 2 3 2 3 2 1 1 1/3 7/9 1 1 1/3
B 1 1 1 3 2 3 1 2 1 1 1 1 1 1 1
C 2 3 2 3 2 3 2 3 1 3/5 3/4 47/60 1 7/9 43/48
D 3 2 3 2 3 4 2 3 5/6 1 3/4 31/36 1 1 3/4
E 3 3 2 4 3 2 1 2 2/3 1/2 1 13/18 2/3 1/2 1
n
ðkÞ ðkÞ ðkÞ ðkÞ - ðkÞ ðkÞ
λj X ij = X i0 - si = θi X i0 , i = mðk - 1Þ þ 1, . . . , mðkÞ , k = 1, . . . , p
j=1
n ðk - 1Þ ðk - 1Þ
ðkÞ ðk - 1Þ ðk - 1Þ
λj Z gj = Z g0 - sgðk - 1Þ - = θi Z g0 , g = hðk - 2Þ þ 1, . . . , hðk - 1Þ , k = 2, . . . , p
j=1
n
λj Z gj = Z g0 þ sðkÞþ
g = φðkÞ
g Z g0 , g = hðk - 1Þ þ 1, . . . , hðkÞ , k = 1, . . . , p - 1
j=1
n
ðkÞ ðkÞ ðkÞ ðkÞþ ðkÞ ðkÞ
λj Y rj = Y r0 þ sr = φr Y r0 , r = sðk - 1Þ þ 1, . . . , sðkÞ , k = 1, . . . , p
j=1
ð12:23Þ
ðk Þ ðk Þ
Since Division k uses X i and Z ðgk - 1Þ to produce Y ðrkÞ and Z g , its average input
distance parameter is:
ðk - 1Þ
1
mðkÞ ðkÞ
X i0 - si
ðkÞ - hðk - 1Þ Z g0 - sgðk - 1Þ -
þ
ðmðkÞ - mðk - 1Þ Þ þ ðhðk - 1Þ - hðk - 2Þ Þ i = mðk - 1Þ þ1
ðkÞ
X i0 g = hðk - 2Þ þ1
ðk - 1Þ
Z g0
1
mðkÞ
si
ðkÞ - hðk - 1Þ sgðk - 1Þ -
=1- þ
ðmðkÞ - mðk - 1Þ Þ þ ðhðk - 1Þ - hðk - 2Þ Þ ðkÞ
i = mðk - 1Þ þ1 X i0 g = hðk - 2Þ þ1 Z g0
ðk - 1Þ
and the average output distance parameter is:
ðkÞ
1
sðkÞ ðkÞ
Y r0 þ sr
ðkÞþ hðkÞ Z g0 þ sðkÞþ
g
þ
ðkÞ
ðs - s ðk - 1Þ Þ þ ðhðkÞ - hðk - 1Þ Þ r = sðk - 1Þ þ1
ðkÞ
Y r0 g = sðk - 1Þ þ1
ðkÞ
Z g0
1
sðkÞ
sr
ðkÞþ hðkÞ sðkÞþ
g
=1þ þ
ðsðkÞ - sðk - 1Þ Þ þ ðhðkÞ - hðk - 1Þ Þ r = sðk - 1Þ þ1 Y r0
ðkÞ
g = hðk - 1Þ þ1
ðkÞ
Z g0
The division efficiency is thus:

mð1Þ ð 1Þ -
1 si
1-
ð1Þ
mð1Þ i=1 X i0
ð1Þ
E0 =
1
sð1Þ
sðr1Þþ hð1Þ sðg1Þþ
1þ þ
sð1Þ þ hð1Þ ð1Þ
r = 1 Y r0 g=1
ð1Þ
Z g0
1
mð k Þ ðk Þ -
si hð k - 1 Þ sðgk - 1Þ -
1- þ
ðmðkÞ - mðk - 1Þ Þ þ hðk - 1Þ - hðk - 2Þ ðk Þ
i = mðk - 1Þ þ1 X i0 g = hðk - 2Þ þ1
ðk - 1Þ
Z g0
ðk Þ
E0 = , k = 2, . . . , p - 1
1
sðkÞ
sðrkÞþ hðkÞ sðgkÞ -
1þ þ
ðsðkÞ - sðk - 1Þ Þ þ hðkÞ - hðk - 1Þ r = sðk - 1Þ þ1
ðk Þ
Y r0 g = hðk - 1Þ þ1 Z g0
ðk Þ
1
mðpÞ
si
ðpÞ - hðp - 1Þ sðgp - 1Þ -
1- þ
ðmðpÞ - mðp - 1Þ Þ þ hðp - 1Þ - hðp - 2Þ i = mðp - 1Þ þ1 X i0
ðp Þ
g = hðp - 2Þ þ1
ðp - 1Þ
Z g0
ðpÞ
E0 = ðpÞ
1
s
sðrpÞþ
1þ
ðsðpÞ - sðp - 1Þ Þ r = sðp - 1Þ þ1 Y r0
ð pÞ
ð12:24Þ
According to the idea of efficiency aggregation, the system efficiency can be

defined as a weighted average of the division efficiencies, where the weight associ-
ated with a division is the proportion of the average output distance parameter of this
division in the total output distance parameter of all divisions (Kao 2014b). That is,
1 1
sðkÞ
srðkÞþ hðkÞ sðgkÞþ
ω ðk Þ = 1þ þ , k = 1, . . . , p - 1
T ðsðkÞ - sðk - 1Þ Þ þ hðkÞ - hðk - 1Þ r = sðk - 1Þ þ1 Y r0
ðk Þ
g = hðk - 1Þ þ1
ðk Þ
Z g0
sðpÞ
1 1 sðrpÞþ
ωðpÞ = 1 þ ðpÞ ðp - 1Þ ðpÞ
T ðs - s Þ r = sðp - 1Þ þ1 Y r0
ð12:25Þ
where
p-1
1
sðkÞ
sr
ðkÞþ hðkÞ sgðkÞþ
T= 1þ þ
k=1 ðsðkÞ - sðk - 1Þ Þ þ ðhðkÞ - hðk - 1Þ Þ r = sðk - 1Þ þ1 Y r0
ðkÞ
g = hðk - 1Þ þ1
ðkÞ
Z g0
sðpÞ ðpÞþ
1 sr
þ 1 þ ðpÞ ðp - 1Þ :
ðs - s Þ r = sðp - 1Þ þ1 Y ðpÞ
r0
The weighted average of the p division efficiencies based on this set of weights is:
p mð1Þ ð1Þ-
ðkÞ 1 1 si
ωðkÞ E0 = 1-
k=1
T mð1Þ i=1 X i0
ð1Þ
p
1
mðkÞ
si
ðkÞ - hðk - 1Þ sðk
g
-1Þ-
þ 1- þ
k =2 ðmðkÞ - mðk -1Þ Þ þ ðhðk -1Þ -hðk -2Þ Þ i= mðk - 1Þ þ1
ðkÞ
X i0 g= hðk -2Þ þ1
ðk -1Þ
Z g0
where the terms in braces in the numerator are the average distance parameter of the
input factors (including endogenous ones) of a division, and those in the denomina-
tor are that of the output factors (including endogenous ones) of a division. The
efficiency aggregation model under constant returns to scale is then:
mð1Þ ð1Þ -
1 1 si
min: 1- þ
T mð1Þ ð1Þ
i =1 X i0
p
1
mðkÞ ðkÞ-
si hðk - 1Þ sðk
g
-1Þ -
1- þ
k=2 ðmðkÞ - mðk - 1Þ Þ þ ðhðk - 1Þ - hðk - 2Þ Þ i=mðk - 1Þ þ1 X i0
ðkÞ
g=hðk - 2Þ þ1
ðk -1Þ
Z g0
s:t: Equation ð12:23Þ

n n
λj Z gj ≥ λj Z gj , g = hðk - 1Þ þ 1, . .. ,hðkÞ , k = 1, . .. ,p - 1
j =1 j=1
λ, s ≥0 ð12:26Þ
Different forms of the constraints for describing the relationships of the interme-
diate products between two stages have been discussed (Tone and Tsutsui 2010), and
the choice of which ones to use depends on the related assumptions. At optimality,
the weights ω(k) associated with Division k are calculated via Eq. (12.25).
Suppose an efficiency decomposition model is desired. In this case only the slack
variables related to the exogenous inputs and outputs should be considered, and
Model (12.26) is modified to:
ðk Þ -
p 1 mðkÞ si
k=1 1-
ðmðkÞ - mðk - 1Þ Þ i = mðk - 1Þ þ1
X i0
ðk Þ
min:
p 1 s ðk Þ sðrkÞþ ð12:27Þ
k=1 1þ
ðsðkÞ - sðk - 1Þ Þ r = sðk - 1Þ þ1 ðk Þ
Y r0
s:t: Constraints of ð12:26Þ
The objective function is the efficiency of the system, and the division efficiencies
are defined in Eq. (12.24). How to decompose the system efficiency to the division
efficiencies has not been generalized to all structures and is an interesting topic for
further research.
12.6 Reversal Links
The series structures discussed in the preceding sections have only one direction,
from the raw materials toward the final products. In the real world there are also
cases that a division in a later stage may affect the operations of a division in an
earlier stage. For example, Kawaguchi et al. (2014) found that the administration
division of a hospital supplies medical beds to the medical examination division for
medical care, and the medical examination division provides the average revenue per
inpatient for the administration division as a consideration of the number of beds to
be supplied to it. Similarly, in studying the performance of a supply chain, Momeni
et al. (2014) assumed that while there are products flowing from one division to the
next, there are also demands for the products reflecting from one division back to the
preceding division. Another example is wastewater treatment (Yang et al. 2014),
where freshwater is used and then becomes wastewater, and the wastewater is then
processed to be reused as freshwater. In all these cases there are links moving
backward in the multi-stage system.
Since this type of system does not have a general structure, we will use a specific
one, as shown in Fig. 12.6, to discuss this, as systems of other structures can be
modeled in a similar manner. In this system, in addition to the conventional flows of
ðk Þ ðk,aÞ
Z g from one division to the next, there are also backward flows of Bl from
Division k to a previous Division a. Most related studies assume the system has
reached a steady state and use the inputs and outputs of a division to measure its
efficiency. Two types of models have appeared in the literature, ratio-form and
slacks-based, and they are discussed below.
Fig. 12.6 A series structure X 3( 2 ) X 4(3) X 5( 4) , X 6( 4)

with reversal links
X 1(1) , X 2(1) Z1(1) Z 2(2) Z 3(3) Y3(4) , Y4(4)

1 2 3 4
B1(3, 2)
B2(4, 2)
Y1(1) Y2( 2 )
12.6 Reversal Links 315
12.6.1 Ratio-Form Efficiency Measures
The ratio-form model has two types: efficiency aggregation and decomposition.
These two types of model have the same constraints and the same definition of
division efficiency. They differ only in the objective function, in that the former
involves all types of inputs and outputs, while the latter only involves the exogenous
inputs and outputs.
Based on Fig. 12.6, the efficiencies of the four divisions are:
ð1Þ ð1Þ
ð1Þ u1 Y 10 þ w1 Z 10
E0 = ð1Þ ð1Þ
v1 X 10 þ v2 X 20
ð2Þ ð2Þ
ð2Þ u2 Y 20 þ w2 Z 20
E0 = ð2Þ ð1Þ ð3,2Þ ð4,2Þ
v3 X 30 þ w1 Z 10 þ c1 B10 þ c2 B20
ð3Þ ð3,2Þ
ð12:28Þ
ð3Þ w3 Z 30 þ c1 B10
E0 = ð3Þ ð2Þ
v4 X 40 þ w2 Z 20
ð4Þ ð4Þ ð4,2Þ
ð4Þ u3 Y 30 þ u4 Y 40 þ c2 B20
E0 = ð4Þ ð4Þ ð3Þ
v5 X 50 þ v6 X 60 þ w3 Z 30
There are several ways to aggregate the division efficiencies to be the system
efficiency. One way that is able to produce an implicitly linear model is via additive
aggregation, in which the system efficiency is defined as a weighted average of the
division efficiencies, and the weight ω(k) associated with a division is the proportion
of the aggregate input consumed by this division in that consumed by all divisions.
That is,
1 ð1Þ ð1Þ
ωð1Þ = v1 X 10 þ v2 X 20
T
1 ð2Þ ð1Þ ð3,2Þ ð4,2Þ
ωð2Þ = v3 X 30 þ w1 Z 10 þ c1 B10 þ c2 B20
T ð12:29Þ
1 ð3Þ ð2Þ
ωð3Þ = v4 X 40 þ w2 Z 20
T
1 ð4Þ ð4Þ ð3Þ
ωð4Þ = v5 X 50 þ v6 X 60 þ w3 Z 30
T
ð1Þ ð1Þ ð2Þ ð1Þ ð3,2Þ ð4,2Þ
where T = v1 X 10 þ v2 X 20 þ v3 X 30 þ w1 Z 10 þ c1 B10 þ c2 B20 þ
ð3Þ ð2Þ ð4Þ ð4Þ ð3Þ
v4 X 40 þ w2 Z 20 þ v5 X 50 þ v6 X 60 þ w3 Z 30 , which is the total aggregate
input consumed by all four divisions.
The weighted average of the division efficiencies is defined as the system
efficiency. The following efficiency aggregation model is thus obtained:
1 ð1Þ ð1Þ ð2Þ ð2Þ ð3Þ ð3,2Þ

max: u1 Y 10 þ w1 Z 10 þ u2 Y 20 þ w2 Z 20 þ w3 Z 30 þ c1 B10
T
ð4Þ ð4Þ ð4,2Þ
þ u3 Y 30 þ u4 Y 40 þ c2 B20
s:t: u1 Y 1j þ w1 Z 1j - ðv1 X 1j þ v2 X 2j Þ ≤ 0, j = 1, . . . , n
ð2Þ ð2Þ ð2Þ ð1Þ ð3,2Þ ð4,2Þ
u2 Y 2j þ w2 Z 2j - ðv3 X 3j þ w1 Z 1j þ c1 B1j þ c2 B2j Þ ≤ 0, j = 1, . . . , n
ð3Þ ð3,2Þ ð3Þ ð2Þ
w3 Z 3j þ c1 B1j - ðv4 X 4j þ w2 Z 2j Þ ≤ 0, j = 1, . . . , n
ð4Þ ð4Þ ð4,2Þ ð4Þ ð4Þ ð3Þ
u3 Y 3j þ þ
u4 Y 4j c2 B2j - ðv5 X 5j þ þ
v6 X 6j w3 Z 3j Þ ≤ 0, j = 1, . . . , n
ur , vi , wg , cl ≥ ε, r = 1, . . . , 4, i = 1, . . . , 6, g = 1, 2, 3, l = 1, 2
ð12:30Þ
At optimality, the division efficiencies are calculated according to Eq. (12.28),

and the system efficiency is the objective value, which is the weighted average of the
four-division efficiencies. The weights associated with the divisions are calculated
from Eq. (12.29).
The efficiency decomposition model differs from the efficiency aggregation
model only in the objective function, in that the endogenous inputs and outputs
are excluded. The complete model is:

u1 Y 10 þ u2 Y 20 þ u3 Y 30 þ u4 Y 40
max: ð1Þ ð1Þ ð2Þ ð3Þ ð4Þ ð4Þ
v1 X 10 þ v2 X 20 þ v3 X 30 þ v4 X 40 þ v5 X 50 þ v6 X 60 ð12:31Þ
The system efficiency measured from the efficiency decomposition is more

realistic for comparison among DMUs, because it is based on the inputs consumed
and outputs produced by the system. The intermediate products that are produced
and consumed within the system are not visible to other DMUs, and their effects are
already reflected in the final outputs. They should thus be counted only in division-
to-division comparisons among different DMUs.
While formulating the efficiency decomposition model is straightforward, the
process of decomposing the system efficiency into division efficiencies is not so
simple, due to the reversal links, and it requires further research.
12.6.2 Slacks-Based Efficiency Measures
The idea of the slacks-based approach is to define the system efficiency as the
product of the average efficiency of the input factors and that of the output factors,
where the input efficiency is expressed by the excessive inputs used, in input slacks,
and the output efficiency by the insufficient outputs that should have been produced,
in output slacks.
The slack variables corresponding to the input and output factors for the system
shown in Fig. 12.6 can be expressed as:
Division 1
n
λj X ij þ si = X i0 , i = 1, 2
j=1
n
λj Z 1j - s1 = Z 10
j=1
n
λj Y 1j - s1 = Y 10
j=1
Division 2
n
λj X 3j þ s3 = X 30
j=1
n
λj Z 1j þ s1 = Z 10
j=1
n
ð2Þ ð3,2Þ ð3,2Þ - ð3,2Þ
λj B1j þ s1 = B10
j=1
n
ð2Þ ð4,2Þ ð4,2Þ - ð4,2Þ
λj B2j þ s2 = B20
j=1
n
λj Z 2j - s2 = Z 20
j=1
n
λj Y 2j - s2 = Y 20 ð12:32Þ
j=1
Division 3
n
λj X 4j þ s4 = X 40
j=1
n
λj Z 2j þ s2 = Z 20
j=1
n
ð3Þ ð3,2Þ ð3,2Þþ ð3,2Þ
λj B1j - s1 = B10
j=1
n
λj Z 3j - s3 = Z 30
j=1
Division 4
n
λj X ij þ si = X i0 , i = 5, 6
j=1
n
λj Z 3j þ s3 = Z 30
j=1
n
ð4Þ ð4,2Þ ð4,2Þþ ð4,2Þ
λj B2j - s2 = B20
j=1
n
λj Y rj - sr = Y r0 , r = 3, 4
j=1
Based on the slack variables related to the input and output factors, the efficien-
cies of the four divisions are as follows:
ð1Þ - ð1Þ -
1 s1 s
1- þ 2 ð1Þ
2 X ð1Þ X 20
ð1Þ
E0 = 10
ð1Þþ ð1Þþ
1 s1 s
1þ þ 1 ð1Þ
2 Y ð1Þ Z 10
10
ð2Þ - ð1Þ - ð3,2Þ - ð4,2Þ -
1 s3 s s s
1- ð Þ
þ 1 ð1Þ þ 1 ð3,2Þ þ 2 ð4,2Þ
4 X 2
Z 10 B10 B20
ð2Þ
E0 = 30
ð2Þþ ð2Þþ
1 s2 s
1þ ð Þ
þ 2 ð2Þ
2 Y 2
Z 20
20
ð12:33Þ
ð3Þ - ð2Þ -
1 s4 s
1- þ 2 ð2Þ
2 X ð3Þ Z 20
ð3Þ
E0 = 40
ð3Þþ ð3,2Þþ
1 s3 s
1þ þ 1 ð3,2Þ
2 Z ð3Þ B10
30
ð4Þ - ð4Þ - ð3Þ -
1 s5 s s
1- ð Þ
þ 6 ð4Þ þ 3 ð3Þ
3 X 4
X 60 Z 30
ð4Þ
E0 = 50
ð4Þþ ð4Þþ ð4,2Þþ
1 s3 s s
1þ þ 4 ð4Þ þ 2ð4,2Þ
3 Y ð4Þ Y 40 B20
30
Regarding the system efficiency, there are, in general, two ways to express this:
efficiency decomposition and aggregation.
For efficiency decomposition the system efficiency is expressed by the slack
ðk Þ
variables related to the exogenous inputs X i and outputs Y ðrkÞ . The model is:
ð1Þ - ð1Þ - ð2Þ - ð3Þ - ð4Þ - ð4Þ -

1 1 s1 s s3 s4 1 s5 s
4- þ 2 ð1Þ - - - þ 6 ð4Þ
4 2 X ð1Þ X 20 X 30
ð2Þ
X 40
ð3Þ 2 X ð4Þ X 60
10 50
min: ð1Þþ ð2Þþ ð4Þþ ð4Þþ
1 s s 1 s3 s
3 þ 1 ð1Þ þ 2 ð2Þ þ ð4Þ
þ 4 ð4Þ
3 Y 10 Y 20 2 Y 30 Y 40
s:t: Constraints of Equation ð12:32Þ
n n
j=1 j=1
n n
j=1 j=1
n n
j=1 j=1
λ, s ≥ 0
ð12:34Þ
It is important to note that Division 3 does not have an exogenous output, making
the number of divisions in the denominator less than that in the numerator. The
numerator must thus be divided by four and the denominator by three to obtain the
average efficiency per division to make them comparable. The inequality constraints
for the intermediate products can be modified according to different assumptions.
The idea of efficiency decomposition is to find a relationship between the system and
division efficiencies. However, this is not easy for slacks-based models, and further
research is needed.
In the efficiency aggregation model the system efficiency is defined as an
aggregation of the division efficiencies, and there are several ways to carry out this
aggregation, as discussed in the preceding chapter. However, the weighted average
is generally used to make the model (implicitly) linear, with the weight defined as the
proportion of the average output efficiency of the division (in reciprocal form) in the
total output efficiency of all divisions. The denominators of the division efficiencies
in Eq. (12.33) show that the weights are:
ð1Þþ ð1Þþ
1 1 s1 s
ωð1Þ = 1þ ð Þ
þ 1 ð1Þ
T 2 Y 1
Z 10
10
ð2Þþ ð2Þþ
1 1 s2 s
ωð2Þ = 1þ þ 2 ð2Þ
T 2 Y ð2Þ Z 20
20
ð3Þþ
ð12:35Þ
ð3,2Þþ
ð3Þ 1 1 s3 s
ω = 1þ þ 1 ð3,2Þ
T 2 Z ð3Þ B10
30
1 1 s3 s s
ωð4Þ = 1þ þ 4 ð4Þ þ 2 ð4,2Þ
T 3 Y ð4Þ Y 40 B20
30
where
ð1Þþ ð1Þþ ð2Þþ ð2Þþ ð3Þþ ð3,2Þþ

1 s1 s1 1 s2 s2 1 s3 s1
T= 1þ ð1Þ
þ ð1Þ
þ 1þ ð2Þ
þ ð2Þ
þ 1þ ð3Þ
þ ð3,2Þ
þ
2 Y 10 Z 10 2 Y 20 Z 20 2 Z 30 B10
1 s3 s4 s2
1þ ð4Þ
þ ð4Þ
þ ð4,2Þ
:
3 Y 30 Y 40 B20
ðk Þ
Taking the weighted average of ω(k) and E 0 as the system efficiency, the
efficiency aggregation model becomes:
N
min:
D
where
ð1Þ - ð1Þ - ð2Þ - ð1Þ - ð3,2Þ - ð4,2Þ -

1 s1 s 1 s3 s s s
N =4- þ 2 ð1Þ - þ 1 ð1Þ þ 1 ð3,2Þ þ 2 ð4,2Þ
2 X ð1Þ X 20 4 X ð2Þ Z 10 B10 B20
10 30
ð3Þ - ð2Þ - ð4Þ - ð4Þ - ð3Þ -
1 s4 s 1 s5 s s
- þ 2 ð2Þ - þ 6 ð4Þ þ 3 ð3Þ
2 X ð3Þ Z 20 3 X ð4Þ X 60 Z 30
40 50
ð1Þþ ð1Þþ ð2Þþ ð2Þþ ð3Þþ ð3,2Þþ
1 s1 s 1 s2 s 1 s3 s
D=4 þ þ 1 ð1Þ þ þ 2 ð2Þ þ þ 1 ð3,2Þ
2 Y ð1Þ Z 10 2 Y ð2Þ Z 20 2 Z ð3Þ B10
10 20 30
1 s3 s s
þ þ 4 ð4Þ þ 2 ð4,2Þ :
3 Y ð4Þ Y 40 B20
30
At optimality, the system efficiency is the objective value, and the division
efficiencies are calculated according to Eq. (12.33). The weights associated with
the divisions are calculated from Eq. (12.35), which are the most favorable ones for
yielding the highest system efficiency for the DMU.
The general multi-stage system has wide applications in the real world, with a review
presented in Kao (2014c). However, as the issue of network DEA has attracted
considerable attention in recent years, many works have appeared since this review.
Although a series system can have as many divisions as desired, except for
theoretical studies the largest system that has appeared in the literature has only
five divisions. For systems without intermediate products, Park and Park (2009)
applied the conventional system distance measure to measure the performance of
20 cable TV service operation units in South Korea for the period 1999–2001. Kao
and Liu (2014) measured the performance of 22 commercial banks in Taiwan for the
period 2009–2011. Wei et al. (2011) proposed a model to find the minimum of the
sum of the distance parameters associated with all divisions, and showed that the
system is efficient if all divisions are. Medina-Borja and Triantis (2014) used an
independent model to evaluate the fundraising efficiency, capacity building, service
quality, and effectiveness of 960 nonprofit human and social service organizations in
the USA.
For ratio-form efficiency measures, Nouri et al. (2013) identified five divisions in
the production system and used the system efficiency as the objective function to
measure the system and division efficiencies of 15 vegetable oil plants in Iran. Lee
and Johnson (2012) designed a hybrid production system of four divisions, capacity
design, demand generation, operations, and demand consumption, to investigate the
performance of 15 US Airlines. Amirteimoori and Kordrostami (2005), Kordrostami
and Amirteimoori (2005), and Amirteimoori and Shafiei (2006) proposed similar
efficiency aggregation models to measure the performance of a series system. In the
first study, each division consumed exogenous inputs to produce outputs, and a
portion of each output served as an endogenous input for the succeeding division. In
the second study, undesirable inputs and outputs were allowed. In the third study, the
intermediate products of a division, to be used as endogenous inputs by the
succeeding division, were allowed to be different from the outputs.
For distance function and slacks-based efficiency measures, Tone and Sahoo
(2003) modified the system distance measure model to investigate returns to scale
in the presence of indivisibilities. Mallikarjun (2015) took both the input and output
distance parameters into account in a three-stage system to measure the efficiency of
27 US airlines. Matthews (2013) studied the risk management and managerial
efficiencies of 15 banks in China with an SBM model. The system was divided
into three divisions, where non-performing loans were inputs for the third one.
Tsutsui and Goto (2009) used a weighted SBM model to study the performance of
90 vertically integrated electric power companies in the USA. Five levels (divisions)
were considered, namely, generation, transmission, distribution, sales, and general
administration, and the weights were the cost shares. Mirhedayatian et al. (2014)
measured the efficiency of a four-stage supply chain with undesirable and fuzzy data
for 10 Iranian companies producing soft drinks. The slacks-based model of Tone and
Tsutsui (2009) was used to evaluate 17 urban rail transit systems in China (Qin et al.
2014), the national innovation system performance of 41 countries regarding R&D,
diffusion and economic efficiency (Chang 2015), and the operations, services, and
sales efficiencies of 22 international airlines (Li et al. 2015). Troutt et al. (2001)
proposed a value-based model for maximizing the throughput per unit of the input at
the first division. Wei and Chang (2011) proposed a model to optimally design a
DMU’s input and output portfolios, in terms of profit maximization, given the
DMU’s available budget.
The series system with reversal links also has several applications. Yang et al.
(2014) evaluated 29 provincial water recycling and utilization systems in China.
Momeni et al. (2014) proposed a four-stage model, including suppliers, manufac-
turers, distributors, and customers, to study the efficiency of supply chains with
reverse logistics. Kawaguchi et al. (2014) measured the efficiency of 112 municipal
hospitals in Japan from 2007 to 2009, where the administration division supplied
medical beds to the medical examination division for medical care, and the latter
returned the revenue from inpatients to the former as a consideration for the number
of beds to be supplied. Shafiee et al. (2014) used a four-stage approach to evaluate
the performance of 22 supply chains of Iranian food industry. The four stages are
learning and growth, production, customer, and financial, and each stage has some
factors that affect the others.
References
Amirteimoori, A., & Kordrostami, S. (2005). DEA-like models for multi-component performance
measurement. Applied Mathematics and Computation, 163, 735–743.
Amirteimoori, A., & Shafiei, M. (2006). Measuring the efficiency of interdependent decision
making sub-units in DEA. Applied Mathematics and Computation, 173, 847–855.
Chang, C. C. (2015). Influences of knowledge spillover and utilization on the NIS performance: A
multi-stage efficiency perspective. Quality & Quantity, 49, 1945–1967.
Kao, C. (2014a). Efficiency decomposition for general multi-stage systems in data envelopment
Kao, C. (2014c). Network data envelopment analysis: A review. European Journal of Operational
Kao, C., & Liu, S. T. (2014). Multi-period efficiency measurement in data envelopment analysis:
The case of Taiwanese commercial banks. Omega, 47, 90–98.
Kawaguchi, H., Tone, K., & Tsutsui, M. (2014). Estimation of the efficiency of Japanese hospitals
using a dynamic and network data envelopment analysis model. Health Care Management
Science, 17, 101–112.
Kordrostami, S., & Amirteimoori, A. (2005). Undesirable factors in multi-component performance
Lee, C. Y., & Johnson, A. L. (2012). Two-dimensional efficiency decomposition to measure the
demand effect in productivity analysis. European Journal of Operational Research, 216,
584–593.
Li, Y., Wang, Y. Z., & Cui, Q. (2015). Evaluating airline efficiency: An application of virtual
frontier network SBM. Transportation Research Part E- Logistics and Transportation Review,
81, 1–17.
Mallikarjun, S. (2015). Efficiency of US airlines: A strategic operating model. Journal of Air
Transport Management, 43, 46–56.
Matthews, K. (2013). Risk management and managerial efficiency in Chinese banks: A network
DEA framework. Omega, 41, 207–215.
Medina-Borja, A., & Triantis, K. (2014). Modeling social services performances: A four-stage DEA
approach to evaluate fundraising efficiency, capacity building, service quality, and effectiveness
in the nonprofit sector. Annals of Operations Research, 221, 285–307.
Mirhedayatian, S. M., Azadi, M., & Saen, R. F. (2014). A novel network data envelopment analysis
model for evaluating green supply chain management. International Journal of Production
Economics, 147, 544–554.
Momeni, E., Tavana, M., Mirzagoltabar, H., & Mirhedayatian, S. M. (2014). A new fuzzy network
slacks-based DEA model for evaluating performance of supply chains with reverse logistics.
Journal of Intelligent & Fuzzy Systems, 27, 793–804.
Nouri, J., Lotfi, F. H., Borgheipour, H., Atabi, F., Sadeghzadeh, S. M., & Moghaddas, Z. (2013). An
analysis of the implementation of energy efficiency measures in the vegetable oil industry of
Iran: A data envelopment analysis approach. Journal of Cleaner Production, 52, 84–93.
Park, K. S., & Park, K. (2009). Measurement of multiperiod aggregative efficiency. European
Qin, F., Zhang, X., & Zhou, Q. (2014). Evaluating the impact of organizational patterns on the
efficiency of urban rail transit systems in China. Journal of Transport Geography, 40, 89–99.
Shafiee, M., Lotfi, F. H., & Saleh, H. (2014). Supply chain performance evaluation with data
envelopment analysis and balanced scorecard approach. Applied Mathematical Modeling, 38,
5092–5112.
References 323
Tone, K., & Sahoo, B. K. (2003). Scale indivisibilities and production function in data envelopment
analysis. International Journal of Production Economics, 84, 165–192.
3–4.
Troutt, M. D., Ambrose, P. J., & Chan, C. K. (2001). Optimal throughput for multistage input-
output processes. International Journal of Operations & Production Management, 21,
148–158.
Tsutsui, M., & Goto, M. (2009). A multi-division efficiency evaluation of U.S. electric power
companies using a weighted slacks-based measure. Socio-Economic Planning Sciences, 43,
201–208.
Wei, Q. L., & Chang, T. S. (2011). Optimal system design series-network DEA models. Journal of
Wei, Q. L., Yan, H., & Pang, L. Y. (2011). Composite network data envelopment analysis model.
International Journal of Information Technology & Decision Making, 10, 613–633.
Yang, F., Du, F., Liang, L., & Yang, Z. (2014). Forecasting the production abilities of recycling
systems: A DEA based approach. Journal of Applied Mathematics, 2014, 961468. https://doi.
org/10.1155/2014/961468
Chapter 13
Parallel Systems
There are two basic structures in network analysis that are the basis for studying
general network structures, series and parallel. For the former, the divisions of a
system are arranged in sequence, one after another; in that the outputs of one division
are the inputs of the next. In general, a division can start its operation only after its
preceding divisions have finished their work. For the latter, all divisions of a system
appear in parallel, in that every division operates independently at the same time,
without affecting each other. The preceding chapter introduced the series structure,
and this chapter will discuss the parallel structure.
Parallel systems can be classified into two types according to the function of the
divisions, multi-component and multi-function. The former is composed of several
divisions of the same function, each using the same inputs to produce the same
outputs, and each DMU need not have the same number of divisions. For example, a
power company usually owns several power plants operating independently. Each
power company has a different number of power plants, but they consume the same
inputs to produce the same outputs. Section 13.1 discusses the efficiency measure-
ment for this type of system. The multi-function system, which will be discussed in
Sect. 13.2, has the same number of divisions for every DMU, in that each division
performs a specific function. For example an international hotel usually has three
divisions in charge of rooms, restaurants, and shops, for the functions of accommo-
dation, dining, and shopping, respectively. Another example is the teaching and
research divisions of universities. One difference between these two examples is that
each division of the former has its own inputs while that of the latter may share some
joint inputs. In the international hotel example, each division has its own inputs, and
they are quite independent in performing the three functions. Many of the inputs for
the teaching and research divisions are in common; they share professors, equip-
ment, expenditures, and so on. Section 13.3 discusses the modeling of shared input.
An extension of the parallel structure is hierarchical structures. When some of the
divisions in a parallel system also have a parallel structure composed of a number of
https://doi.org/10.1007/978-3-031-27593-7_13
326 13 Parallel Systems
sub-divisions at a lower level, then we have a hierarchical system. In short, a

hierarchical system is one with two or more levels of parallel structures.
The way that a hierarchical system operates is that the top management of the
organization allocates the inputs to the divisions at the first level. The divisions at the
first level then allocate the inputs they receive from the headquarters to their
subordinate divisions at the second level. This input allocation process is continued
to the divisions at the bottom of the structure of the organization. The outputs of the
divisions at the lowest level are what they produce. For divisions of other levels,
their outputs are those of their subordinate divisions. The top management is
concerned with the performance of the divisions at the first level, and the heads of
the divisions at this level are concerned with the performance of their subordinate
divisions at the second level. This situation continues to the lowest level, where the
heads of the divisions are concerned with the performance of their own operations.
Section 13.4 discusses how to measure the efficiencies of the divisions at different
levels for systems with hierarchical structures. It is interesting to note that the model
for measuring the efficiencies of a hierarchical system resembles that for measuring
the efficiencies of a parallel system composed of the divisions at the bottom of the
hierarchical structure.
Many applications of the parallel DEA model have been reported, and they will
be briefly described in Sect. 13.5.
13.1 Multi-component Systems
Kao (2009b) studied a problem of efficiency measurement for the forest districts in
Taiwan, where each district uses the inputs of land, labor, forest stocks, and
operating expenses to generate timber, recreational visitors, and a soil conservation
effect. Each district is divided into a number of working circles performing the same
functions. In this example, the district is the DMU, and the working circles are
divisions. We will use the idea from Kao (2009b) to discuss the efficiency measure-
ment of multi-component systems.
As depicted in Fig. 13.1, DMU j has pj divisions, and each division applies the
same inputs Xi, i = 1, . . ., m, to produce the same outputs Yr, r = 1, . . ., s. The total
inputs consumed and total outputs produced by the pj divisions of DMU j are
p ðk Þ p ðk Þ
X ij = k j= 1 X ij , i = 1, . . ., m, and Y rj = k j= 1 Y rj , r = 1, . . ., s, respectively. To
measure the relative efficiency of a DMU in a network setting, Kao (2009a)
proposed a relational model that requires the same factor to have the same multiplier
to maintain the relation of the divisions in the system. Based on this idea, while
taking the operations of all divisions into account, the model for measuring the
efficiency of this system under constant returns to scale is:
13.1 Multi-component Systems 327
Fig. 13.1 Structure of the DMU j

multi-component system
X ij(1) Yrj(1)
1
i=1,..., m r =1,…, s
X ij( 2) Yrj(2)
2
i=1,..., m r =1,…, s
.
.
.
X ij( p j )
Yrj( p j )
pj
i=1,..., m r =1,…, s
s
r = 1 ur Y r0
E 0 = max: m
vi X i0
i=1
s m
ð13:1Þ
ðkÞ ðkÞ
s:t: ur Y rj - vi X ij ≤ 0, k = 1, . . . , pj , j = 1, . . . , n
r=1 i=1
ur , vi ≥ ε, 8r, i
The objective function is a linear fractional, which can be linearized by assigning

1 to the denominator to be a constraint, and then using the numerator as the new
objective function. The constraints of this model indicate that the production frontier
is constructed by the divisions of all DMUs.
At optimality, the system and division efficiencies are calculated from the objec-
tive function and constraints as follows:
s
r =1 ur Y r0
E0 = m
i=1 vi X i0
s ðk Þ ð13:2Þ
ðk Þ r =1 ur Y r0
E0 = k =1, ...,p0
m ðk Þ
i=1 vi X i0
Since the divisions operate independently, without intermediate products

connecting each other, the weighted average of the division efficiencies will equal
the system efficiency. This can be verified from two aspects, input and output. From
the input side, let the weight associated with Division k be the proportion of the
aggregate input consumed by this division in that consumed by all divisions, that is,
m ðk Þ
i = 1 vi X i0
ωðkÞ = m ð13:3Þ
i = 1 vi X i0
with k 0= 1 ωðkÞ = 1 and ω(k) ≥ 0, k = 1, . . ., p0. The weighted average of the p0

p
division efficiencies can be derived as:
p0 p0 m ðk Þ s ðk Þ p0 s ðk Þ
ðk Þ i = 1 vi X i0 r = 1 ur Y r0 r = 1 ur Y r0
ωðkÞ E 0 = m × = m
m ðk Þ
k=1 k=1 i = 1 vi X i0 i = 1 vi X i0 k=1 i = 1 vi X i0
s
r = 1 ur Y r0
= m
i = 1 vi X i0
which indicates that the system efficiency can be decomposed into a weighted
arithmetic average of the division efficiencies. Since the weighted average of the
division efficiencies is equal to the system efficiency, the efficiency decomposition
model for the multi-component system is the same as the efficiency aggregation one.
Model (13.1) is an input-oriented model. The output-oriented model has an
objective function of 1/E0= min. m i = 1 vi X i0 =
s
r = 1 ur Y r0 and the same constraints
as Model (13.1). If we define the weight of Division k to be the proportion of the
aggregate output produced by this division in that produced by all divisions, that is,
s ðk Þ
r = 1 ur Y r0
ωðkÞ = s
r = 1 ur Y r0
then the weighted average of the p0 division efficiencies, in reciprocal form,

becomes:
p0 p0 s ðk Þ m ðk Þ m
1 r = 1 ur Y r0 i = 1 vi X i0 i = 1 vi X i0 1
ωðkÞ ðk Þ
= s
× = s
=
E0 r = 1 ur Y r0
s ðk Þ r = 1 ur Y r0
E0
k=1 k=1 r = 1 ur Y r0
which indicates that the system efficiency is a weighted harmonic average of the
division efficiencies. Again, since the system efficiency can be decomposed into the
division efficiencies via a weighted harmonic average relationship, and the weighted
harmonic average of the division efficiencies produces the same system efficiency as
that measured from the inputs and outputs of the system, the efficiency decomposi-
tion and aggregation models are the same.
If the system is treated as a black box, then the constraints of the corresponding
model for measuring the system efficiency will be the difference between the
aggregate output and aggregate input of the system, sr = 1 ur Y rj - m
i = 1 vi X ij ≤ 0,
j = 1, . . ., n, which is the sum of the pj constraints of a DMU in Model (13.1),
pj s ðk Þ m ðk Þ
k=1 r = 1 ur Y rj - i = 1 vi X ij ≤ 0, j = 1, . . ., n. This implies that the black-
box, or CCR, model is less stringent than the parallel model in measuring efficiency.
13.1 Multi-component Systems 329
Table 13.1 Data and efficiency measures for the multi-component example
DMU X Y CCR efficiency Parallel efficiency (ω(k))
A 5 9 1 9/10
a1 2 4 1 (2/5)
a2 3 5 5/6 (3/5)
B 8 10 25/36 5/8
b1 1 1 1/2 (1/8)
b2 3 3 1/2 (3/8)
b3 4 6 3/4 (4/8)
The system efficiency measured from the CCR model will thus be greater than or
equal to that measured from Model (13.1).
The property that the system efficiency is a weighted average of the division
efficiencies indicates that a system is efficient if and only if all its divisions are
efficient. It is thus possible that none of the DMUs is efficient, which is different
from the black-box measurement, in that at least one DMU will be efficient. Another
point to be noted is that theoretically all divisions from different DMUs are compa-
rable, because they use the same inputs to produce the same outputs, and they
construct the frontier together. However, due to possibly multiple solutions, the
division efficiencies calculated from Eq. (13.2) may not be compared with each
n
other. If a comparison is desired, then one must treat all j = 1 pj divisions as
independent DMUs, and apply the CCR model to calculate their efficiencies. The
efficiencies thus measured are then comparable.
Consider a very simple example of two DMUs, A and B, using one input X to
produce one output Y, where A has two divisions, a1 and a2, and B has three
divisions, b1, b2, and b3, with the data shown in columns two and three of
Table 13.1. When the conventional black-box model is used, the efficiencies of
1 and 25/36 are obtained for DMUs A and B, as shown in column four of Table 13.1
under the heading of “CCR efficiency.” By applying Model (13.1), together with
Eq. (13.2), the system and division efficiencies are calculated, with the results shown
in the second-to-last column. The values in parentheses in the last column are the
weights associated with each division calculated from Eq. (13.3).
Figure 13.2 shows that the production frontier constructed from the five divisions,
a1, a2, b1, b2, and b3, is ray Oa1, based on which the division and system efficiencies
are measured. For example, the efficiency of Division b3 is X / X b3 , which is 3/4,
b3
and the efficiency of DMU A is X / XA, which is 9/10. Note that the division
A
efficiencies calculated from Eq. (13.2) in this example are the same as those
measured from the CCR model by treating each division as an independent DMU.
The results in Table 13.1 show that, as expected, the system efficiencies measured
from the parallel model are less than those measured from the black-box model.
Moreover, neither of the two DMUs is efficient. It is also easy to verify that the
weighted average of the division efficiencies is equal to the system efficiency. For
Fig. 13.2 Frontier and Y

efficiency measurement of
the multi-component B
example 10
Â A
b̂3 b3
6
a2
4 a1
b2
2
b1
X b̂
3
X b X Aˆ X A
3
O 2 4 6 8 10 X
DMU A it is (2/5) × (1) + (3/5) × (5/6) = 9/10, and for DMU B it is

(1/8) × (1/2) + (3/8) × (1/2) + (4/8) × (3/4) = 5/8..
Model (13.1) is of ratio form. By setting the denominator of the objective function
to 1 and keeping the numerator as the objective function to transform into a linear
program, its dual, which is of envelopment form, can be formulated as:
min: θ
n pj
ðk Þ ðk Þ ð13:4Þ
s:t: λj X ij ≤ θXi0 , i = 1, . . . , m
j=1 k=1
n pj
ðk Þ ðk Þ
λj Y rj ≥ Y r0 , r = 1, . . . , s
j=1 k=1
ðk Þ
λj ≥ 0, k = 1, . . . , pj , j = 1, . . . , n
Here the non-Archimedean number ε is omitted for simplicity of expression. This

model is able to show the targets for the inefficient DMUs to follow in order to
become efficient. Specifically, we have,
13.2 Multi-function Systems 331
n pj
ðk Þ ðk Þ
X i0 = λj X ij , i = 1, . . . , m
j=1 k=1
n pj
ðk Þ ðk Þ
Y r0 = λj Y rj , r = 1, . . . , s
j=1 k=1
where only the total inputs X i0 and total outputs Y r0 of the DMU are shown. This
model is not able to separate the total value into individual divisions. If the targets for
individual divisions are desired, then the inputs and outputs of every division must
be treated as different factors to formulate a different envelopment model.
A final remark to make for this section is that, structurally, the multi-component
system resembles the multi-stage system if we treat each division as a stage.
Measuring the efficiencies of the DMU and divisions of a multi-component system
is thus the same as measuring those of the DMU and stages of a multi-stage system.
13.2 Multi-function Systems
The case discussed in the preceding section refers to systems with different numbers
of divisions of the same function that can be compared not only within the same
DMU, but also among different DMUs. Another case that is more common in the
real world is systems composed of the same number of divisions, and each division
has specific functions. Since the divisions of a DMU are not homogenous, they
cannot be compared with each other. The divisions of different DMUs of the same
function, however, are comparable. For example, the post office in Taiwan has two
major services, postal services, and banking. These two services have different
inputs and outputs and are thus not comparable within the same branch. The same
postal or banking services of different branches are, however, comparable.
Figure 13.3 shows the structure of the multi-function system, which is similar to
that of the multi-component system depicted in Fig. 13.1. The differences are that the
number of divisions in the multi-function system is the same for all DMUs, and the
inputs and outputs of different divisions of the same DMU are basically different.
Note that for the same input, e.g., employees, if they cannot be shared among
divisions then they are considered as different inputs. This is quite normal in the
real world. For the post office example, the staff working in the postal division will
not help those in the banking division. Referring to Fig. 13.3, Division k applies
ðk Þ
m(k) - m(k - 1) inputs X ij , i = m(k - 1) + 1, . . ., m(k), to produce s(k) - s(k - 1) outputs
ðk Þ
Y rj , r = s(k - 1) + 1, . . ., s(k), where m(0) = s(0) = 0 and m( p) = m, s( p) = s. Since all
ðk Þ ðk Þ
inputs and outputs are considered different, the superscript (k) in X ij and Y rj is not
necessary. However, it is still used to better identify the division of concern.
The efficiency of the kth division of a DMU, as compared to those of other DMUs
under constant returns to scale, can be measured via the following model:
Fig. 13.3 Structure of the DMU j

multi-function system
X ij(1) Yrj(1)
1
i = m ( 0 ) + 1, ..., m (1) r = s ( 0) + 1, ..., s (1)
X ij( 2) Yrj( 2)
2
i = m (1) + 1, ..., m ( 2 ) r = s (1) + 1, ..., s ( 2)
.
.
.
X ij( p) Yrj( p )
p
i = m ( p –1) + 1, ..., m ( p ) r = s ( p –1) + 1, ..., s ( p )
sðkÞ ðkÞ
ðkÞ r = sðk - 1Þ þ1 ur Y r0
E0 = max: mðkÞ ðkÞ
i = mðk - 1Þ þ1 vi X i0
sðkÞ mðkÞ ð13:5Þ
ðkÞ ðkÞ
s:t: ur Y rj - vi X ij ≤ 0, j = 1, . . . , n
r = sðk - 1Þ þ1 i = mðk - 1Þ þ1
ur , vi ≥ ε, 8r, i
There are n constraints, one for each DMU. When all p divisions are considered
together, the model becomes:
s
r = 1 ur Y r0
E 0 = max: m
i = 1 vi X i0
sðkÞ mðkÞ
ðkÞ ðkÞ
s:t: ur Y rj - vi X ij ≤ 0, k = 1, . . . , p, j = 1, . . . , n
r = sðk - 1Þ þ1 i = mðk - 1Þ þ1
ur , vi ≥ ε, 8r, i
ð13:6Þ
There are p × n constraints corresponding to the p divisions in each of the

n DMUs. Models (13.5) and (13.6) look very similar, except the former involves
only the inputs and outputs of Division k, while the latter involves those of all
divisions.
When a set of optimal solutions (u*, v*) is obtained, the system and division
efficiencies, based on the objective function and constraints, are calculated as:
13.2 Multi-function Systems 333
s
r = 1 ur Y r0
E0 = m
i = 1 vi X i0
sðkÞ ðk Þ
ð13:7Þ
ðk Þ r = sðk - 1Þ þ1 ur Y r0
E0 = , k = 1, . . . , p
m ð k Þ
ðk Þ
i = mðk - 1Þ þ1 vi X i0
If we define the weight for Division k as the proportion of the aggregate input
consumed by this division in that consumed by all divisions, that is,
mðkÞ ðk Þ
ðk Þ i = mðk - 1Þ þ1 vi X i0
ω = m , k = 1, . . . , p ð13:8Þ
i = 1 vi X i0
then the system efficiency can be decomposed into a weighted average of the
p division efficiencies:
s p s ðk Þ ðk Þ p
r = 1 ur Y r0 r = sðk - 1Þ þ1 ur Y r0
m = m =
i = 1 vi X i0 k=1 i = 1 vi X i0 k=1
ðk Þ ðk Þ
m ðk Þ s ðk Þ p
i = mðk - 1Þ þ1 vi X i0 r = sðk - 1Þ þ1 ur Y r0 ðk Þ
m × = ωð k Þ E 0
mðkÞ ðk Þ
i = 1 vi X i0 i = mðk - 1Þ þ1 vi X i0 k=1
Conversely, if we define the system efficiency as the weighted average of the

p division efficiencies, with the weights defined in Eq. (13.8), then the system
efficiency becomes:
p p mðkÞ ðkÞ sðkÞ ðkÞ

ðkÞ i = mðk - 1Þ þ1 vi X i0 r = sðk - 1Þ þ1 ur Y r0
ωðkÞ E 0 = m × ðkÞ
ðkÞ
i = 1 vi X i0
m
k=1 k=1 i = mðk - 1Þ þ1 vi X i0
p sðkÞ ðkÞ s
r = sðk - 1Þ þ1 ur Y r0 r = 1 ur Y r0
= m = m
k=1 i = 1 vi X i0 i = 1 vi X i0
which is the same as that defined in the objective function of the efficiency decom-
position model (13.6). Model (13.6) is thus also an efficiency aggregation model.
Similar to the case of the series system, there usually exist multiple solutions,
which make the efficiencies of different DMUs incomparable. The two-step
approach discussed in the preceding chapter can be used to compare the efficiency
of Division k of different DMUs when alternative solutions exist. That is, one uses
Model (13.6) to find the system efficiency E0 in the first step and then measures the
sðkÞ ðk Þ mðkÞ ðk Þ
efficiency of Division k by using r = sðk - 1Þ þ1 ur Y r0 / i = mðk - 1Þ þ1 vi X i0 as the
objective function, while requiring the system efficiency to equal E0, expressed as
s m
r = 1 ur Y r0 = E 0 i = 1 vi X i0 , in the second step.
Model (13.6) is of ratio form, which is not able to show the targets for inefficient
DMUs to follow to become efficient, although it is able to calculate the division
efficiencies. An envelopment model is needed in order to find a set of targets. The
dual of Model (13.6), which is of envelopment form, can be formulated as:
min: θ
n
ðkÞ ðkÞ ðkÞ
j=1
n
ðkÞ ðkÞ ðkÞ
ð13:9Þ
j=1
ðkÞ
λj ≥ 0, k = 1, . . . , p, j = 1, . . . , n
In this model, the non-Archimedean number is also omitted. Based on the

constraints of this model, the targets are obtained as:
n
^ i0 = ðkÞ ðkÞ
X λj X ij , i = mðk - 1Þ þ 1, . . . , mðkÞ , k = 1, . . . , p
j=1
n
ðkÞ ðkÞ
Y^ r0 = λj Y rj , r = sðk - 1Þ þ 1, . . . , sðkÞ , k = 1, . . . , p
j=1
Different from the multi-component system, where only the targets for the DMU
are obtainable, here the targets for all divisions can be derived. However, it should be
noted that the above equations are based on the assumption that the solution to
Model (13.6) is unique. If there exist multiple solutions, then the two-step approach
must be applied, and the dual of the second-step model must be formulated to be able
to find the correct targets.
13.3 Shared Input
For systems composed of several divisions performing different functions, there are
usually joint inputs being shared by all, or a number of, divisions, in addition to their
specific inputs. For example, the professors in a university have the responsibilities
of teaching and research. In the terminology of DEA, the input professor is jointly
used by the divisions of teaching and research, which, together with other division-
specific inputs, produce different outputs. Another type of shared input is when the
input can be used by only one division at a time. Examples of this type include the
land for producing different crops and the floor area of a store for displaying different
items. If an area is used by one division, then it cannot be used by another, and the
total area for different divisions is fixed. Most parallel systems in the real world have
shared input, and most studies in parallel systems also concentrate on these.
DMU j
αl(1)
X ij(1) Yrj(1)
1
i = m( 0) + 1, ..., m(1) r = s ( 0 ) + 1, ..., s (1)
αl(2)
X ljS X ij( 2) Yrj( 2)

2
l=1,..., i = m(1) + 1, ..., m( 2) r = s (1) + 1, ..., s ( 2 )
.
.
.
α l( p )
X ij( p) Yrj( p )
p
i = m( p –1) + 1, ..., m( p ) r = s ( p – 1) + 1, ..., s ( p )
Fig. 13.4 Structure of the parallel system with shared input
Figure 13.4 is the structure of a parallel system with shared input, in that every
ðk Þ
division k consumes the division-specific inputs X i , i = m(k - 1) + 1, . . ., m(k), and a
ðk Þ
proportion αl of the lth shared input X Sl , l = 1, . . ., q, which is shared by all
divisions to produce the specific outputs Y ðrkÞ , r = s(k - 1) + 1, . . ., s(k). The divisions
can also produce shared output, such as the reputation of a DMU, which is an
aggregation effect of all divisions. Since the way of handling shared outputs is
similar to that of handling shared inputs, the former will not be discussed. Several
approaches for measuring the efficiency of parallel systems with shared input have
been proposed, including ratio-form, distance function, and slacks-based, and this
section examines these three types.
13.3.1 Ratio-form Efficiency Measures
The system efficiency can be measured based on the total input it consumes and total
output it produces while taking the operations of all divisions into consideration.
Following the structure of the parallel system in Fig. 13.4, the ratio-form model
under constant returns to scale is:
s
E0 = max: ur Y r0
r =1
m q
s:t: vi X i0 þ t l X Sl0 = 1
i =1 l=1
sðkÞ mðkÞ q
ðkÞ ðkÞ ðkÞ
ur Y rj - vi X ij þ t l αl X Slj ≤ 0, k = 1, ... ,p, j= 1, ..., n
r = sðk - 1Þ þ1 i= mðk - 1Þ þ1 l=1
t l ,ur ,vi ≥ ε, 8l,r, i

ð13:10Þ
Note that the first constraint is the result of a linear transformation of the original
s m q
r = 1 ur Y r0 = i = 1 vi X i0 þ l = 1 t l X l0 : After an optimal
S
objective function of
solution (t , u , v ) is obtained, the system and division efficiencies are calculated as:
* * *
s
r = 1 ur Y r0
E0 = m q S
i = 1 vi X i0 þ l = 1 t l X l0
sðkÞ ðk Þ
ð13:11Þ
E0 = , k = 1, . . . , p
mðkÞ ðk Þ q ðk Þ S
i = mðk - 1Þ þ1 vi X i0 þ l = 1 t l αl X l0
As discussed in the preceding section, if we define the weight corresponding to

Division k as the proportion of the aggregate input consumed by this division in that
consumed by all divisions, that is,
m ðk Þ ðk Þ q ðk Þ S
ωðkÞ = m q S
, k = 1, . . . , p ð13:12Þ
i = 1 vi X i0 þ l = 1 t l X l0
then the weighted average of the division efficiencies is:
mðkÞ ðk Þ q ðk Þ S s ðk Þ ðk Þ
p p
ωðkÞ E 0 =
× ð Þ ðk Þ S
m q
i = 1 vi X i0 þ
S mðkÞ k q
k=1 k=1 l = 1 t l X l0 i = mðk - 1Þ þ1 vi X i0 þ l = 1 t l αl X l0
p sðkÞ ðk Þ s
r = sðk - 1Þ þ1 ur Y r0 r = 1 ur Y r0
= m q S
= m X þ q S
k=1 i = 1 vi X i0 þ l = 1 t l X l0 v
i = 1 i i0 l = 1 t l X l0
which is equal to the system efficiency E0. The efficiency decomposition and
efficiency aggregation models are thus also the same for this type of parallel system.
In many cases how an input is shared by all divisions is not clear. In the university
professors example, how much time a professor spends on teaching and research is
usually not clear, and it is only known that the time lies in a reasonable range of [Ll,
Ul]. In this case several approaches can be applied to measure efficiencies, such as
treating the time as interval or fuzzy data, and applying the methods used for
handling these, as discussed in Chap. 7. The method that is most often used is to
Fig. 13.5 A two-division α

parallel system with shared
input
X1 1 Y1
S
X
X2 2 Y2
1– α
Table 13.2 Data and the ratio-form efficiencies for the shared-input example
ð1Þ ð2Þ
DMU X1 X2 XS Y1 Y2 E0 E0 (ω(1)) E0 (ω(2))
A 3 2 4 3 2 0.6666327 0.6666567 (0.99964) 0.6 (0.00036)
α* = 2/3: 0.6666333 0.6666555 (0.99967) 0.6 (0.00033)
B 4 5 5 6 4 0.999732 1 (0.99930) 0.6171 (0.00070)
α* = 0.7974: 0.999733 1 (0.99933) 0.5986 (0.00067)
C 6 4 6 5 3 0.6565 0.69431 (0.6) 0.59985 (0.4)
α* = 0.8: 0.6754 0.69434 (0.7998) 0.59970 (0.2002)
D 6 3 6 4 5 0.9995 0.4861 (0.00096) 1 (0.99904)
α* = 0.2: 0.99956 0.4762 (0.00084) 1 (0.99916)
E 4 6 8 5 5 0.83296 0.833183 (0.99908) 0.5870 (0.00092)
α* = 0.2: 0.83303 0.833283 (0.99876) 0.6290 (0.00124)
ðk Þ
treat the proportion αl as a variable, and search for the most favorable value that
ðk Þ
will yield the highest efficiency for the system. Based on this idea, αl , l = 1, . . .,
q, k = 1, . . ., p, in Model (13.10) are treated as variables, with the following
constraints added:
ðkÞ ðkÞ ðkÞ

Ll ≤ αl ≤ U l , l = 1, . . . , q, k = 1, . . . , p
p
ðkÞ
ð13:13Þ
αl = 1, l = 1, . . . , q
k=1
After an optimal solution is obtained, the system and division efficiencies are
calculated via equations similar to those in Eq. (13.11).
Consider a simple example of five DMUs, denoted as A, B, C, D, and E,
composed of two divisions, with the structure shown in Fig. 13.5, where Division
1 applies input X1 and proportion α of the shared input XS to produce output Y1, and
Division 2 applies input X2 and proportion 1-α of the shared input XS to produce
output Y2. Columns two to six of Table 13.2 show the data for this example. Let
α = 0.6 and ε = 0.0001. By applying Model (13.10) the system efficiencies are
obtained, as shown in column seven of Table 13.2 (in the first row of every DMU).
The division efficiencies and the associated weights are calculated via Eqs. (13.11)
and (13.12), with the results shown in the last four columns (in the first row of every
DMU), where the values in parentheses are the weights. In order to show the
difference, the values are expressed in different numbers of digits for each DMU.
It is easy to verify that the system efficiency E0 is the weighted average of the two
ðk Þ
division efficiencies E0 . For example, the system efficiency of DMU A is
0.6666327, and the weighted average of its two division efficiencies is
0.99964 × 0.6666567 + 0.00036 × 0.6 = 0.6666327, which are the same.
In this example, the proportion α is fixed at 0.6. If it is treated as a variable,
allowing for each DMU to select the most favorable value to measure the efficiency,
then Model (13.10), added with the constraints in (13.13), is applied. Suppose the
bounds L and U are set to 0.2 and 0.8, respectively. The last five columns of
Table 13.2 (in the second row of each DMU) show the results. Comparing the
results obtained from fixing α at 0.6 and varying it in the range of 0.2 and 0.8 it is
noted that the system efficiency of the latter is greater than the former for every
DMU, although the difference is very small. This is simply because allowing α to
vary increases the weight of the division with greater efficiency. For example,
Division 1 of DMU A has a higher efficiency than Division 2, its weight thus
increases from 0.99964 to 0.99967, with α = 2/3.
ðk Þ
Allowing the proportion αl to vary makes the model nonlinear, and thus it is
difficult to obtain solutions of high precision. Fortunately, in most cases, the
nonlinear model can be linearized by applying an appropriate variable substitution
technique.
13.3.2 Distance Function Efficiency Measures
Distance function efficiency measures have several forms, and here we will intro-
duce three of them, those attaching the distance parameters to either the input or
output factors, both the input and output factors, and the directional vector. Other
forms will only be briefly described in Sect. 13.4, Supplementary Literature.
13.3.2.1 System Parameter
Model (13.10) is of ratio form from the input side, whose dual is an input-oriented
distance function model, which can be formulated as:
min: θ
n
ðkÞ ðkÞ ðkÞ
j=1
p n
ðkÞ ðkÞ
λj αl X Slj ≤ θX Sl0 , l = 1, . . . , q
k=1 j=1
n
ðkÞ ðkÞ ðkÞ
j=1
ðkÞ
λj ≥ 0, k = 1, . . . , p, j = 1, . . . , n
ð13:14Þ
In this model the non-Archimedean number ε is omitted for simplicity of expres-

sion. The merit of this type of model is that it is able to provide a set of targets for
inefficient DMUs to follow in order to become efficient. Based on the constraints, the
targets for the input and output factors are:
n
ðk Þ ðk Þ
X i0 = λj X ij , i = 1, . . . , m
j=1
p n
S ðk Þ ðk Þ
X l0 = λj αl X Slj , l = 1, . . . , q ð13:15Þ
k=1 j=1
n
ðk Þ ðk Þ
Y r0 = λj Y rj , r = 1, . . . , s
j=1
where the subscript i running from 1 to m is equivalent to running from m(k - 1) + 1 to

m(k) for k = 1, . . ., p, and the subscript r running from 1 to s is equivalent to running
from s(k - 1) + 1 to s(k) for k = 1, . . ., p, expressed in Model (13.14). It should be
noted that the target for the shared input X Sl is the total amount for the whole DMU.
ðk Þ
This does not imply that αl X Sl0 is the target for Division k. If the target value of X Sl
for Division k is desired, then the current amount of X Sl being used by Division
k must be known, and the target value can be obtained as follows.
Sðk Þ
Let X lj be the amount of X Sl being used by Division k of DMU j. The (unknown)
Sðk Þ
target amount xl is obtained via the following output-oriented model:
max: φ
n
ðkÞ ðkÞ ðkÞ
s:t: λj X ij ≤ X i0 , i = mðk - 1Þ þ 1, . . . , mðkÞ , k = 1, . . . , p
j=1
n
ðkÞ SðkÞ SðkÞ
λj X lj ≤ xl , l = 1, . . . , q, k = 1, . . . , p
j=1
n
ðkÞ ðkÞ ðkÞ
λj Y rj ≥ φY r0 , r = sðk - 1Þ þ 1, . . . , sðkÞ , k = 1, . . . , p
j=1
p
SðkÞ
xl ≤ X Sl0 , l = 1, . . . , q
k=1
ðkÞ SðkÞ
λj , xl ≥ 0, k = 1, . . . , p, l = 1, . . . , q, j = 1, . . . , n
ð13:16Þ
Sð k Þ
The constraint pk = 1 xl ≤ X Sl0 requires that the total amount of the shared input
X Slto be used by all p divisions does not exceed the current level of X Sl0 . After the
Sðk Þ
optimal solution is obtained, xl is the target amount of the lth shared input for
Division k. This model thus shows how the amount of the shared input should be
SðkÞ
reallocated. If pk = 1 xl is less than X Sl0 , then an excessive amount of X Sl0 has been
used. The reason for using an output-oriented model is to keep the model linear. If an
Sðk Þ
input model is used, then the distance parameter must be attached to the input xl in
the second constraint of Model (13.16), which results in a nonlinear model.
13.3.2.2 Division Parameters
Model (13.14) attaches the same distance parameter θ to the inputs of all divisions.
This can be relaxed to allow each division to have different parameters, with an
expectation that these parameters represent the efficiency of the associated division.
Let θ(k) be the parameter associated with Division k. Suppose the importance of
the kth division to the system, as perceived by the decision maker, is ω(k), such that
the system efficiency is the weighted average of θ(k), expressed as θ = pk = 1 ωðkÞ θðkÞ :
We then have the following model to measure the system efficiency:
p
min: ωðkÞ θðkÞ
k=1
n
ðkÞ ðkÞ ðkÞ
s:t: λj X ij ≤ θðkÞ X i0 , i = mðk - 1Þ þ 1, . . . , mðkÞ , k = 1, . . . , p
j=1
p n p
ðkÞ ðkÞ ðkÞ
λj αl X Slj ≤ θðkÞ αl X Sl0 , l = 1, . . . , q
k=1 j=1 k=1
n
ðkÞ ðkÞ ðkÞ
j=1
ðkÞ
λj ≥ 0, k = 1, . . . , p, j = 1, . . . , n
ð13:17Þ
ðk Þ
where the parameters ω(k) and αl are pre-specified positive constants, with
p ðk Þ p ðk Þ
k=1 ω = 1 and α
k=1 l = 1, l = 1, . . ., q. There is a fundamental difference
between this model and the ratio model (13.10). While the system efficiency is a
weighted average of the division efficiencies in both models, the weight associated
with a division in the ratio model is the proportion of the aggregate input consumed
by this division in that consumed by all divisions, which is reflected from the data,
and the weight in the distance function model is subjectively specified by the
decision maker. Anyone applying the ratio model obtains the same results. In
contrast, different people applying the distance function model (13.17) obtain
different results if different weights ω(k) are assigned.
To further investigate the difference between the division efficiencies measured
from Model (13.17) and those measured from Model (13.10), we formulate the dual
of Model (13.17) as follows:
s
E 0 = max: ur Y r0
r =1
mðkÞ q
ðk Þ ðk Þ
s:t: vi X i0 þ tl αl X Sl0 =ωðkÞ , k =1, ...,p
i=mðk -1Þ þ1 l=1
sðkÞ mðkÞ q
ur Y rj - vi X ij þ tl αl X Slj ≤0, k =1, ...,p, j=1, ...,n
r =sðk -1Þ þ1 i=mðk -1Þ þ1 l=1
t l ,ur ,vi ≥ε, 8l,r,i
ð13:18Þ
Comparing this model with Model (13.10), it is noted that only the first constraint
q
is different, in that m i = 1 vi X i0 þ l = 1 t l X l0 = 1 in Model (13.10) is separated into
S
mðkÞ ðk Þ
p constraints of i = mðk - 1Þ þ1 vi X i0 þ l = 1 t l αl X Sl0 = ωðkÞ , k = 1, . . ., p. In Model
q
(13.10) the weight associated with Division k is

Table 13.3 Data and the division efficiencies for the shared-input example
ð1Þ ð2Þ
DMU E0 E0 E0 θ θ(1) θ(2)
A 0.6333233 0.6666466 0.6 0.6333233 0.6666667 0.6
B 0.8998720 1 0.799744 0.899872 1.32 0.48
θ(1), θ(2) ≤ 1: 0.97976 1 0.96
C 0.6220789 0.6942778 0.54988 0.6220789 0.7944444 0.45
D 0.7777111 0.5554222 1 0.7777111 0.5555555 1
E 0.6665967 0.8330334 0.50016 0.6665967 0.8333333 0.5
m ðk Þ q ðk Þ S m q
i = mðk - 1Þ þ1 vi X i0 þ l = 1 t l αl X l0 = i = 1 vi X i0 þ
S
l = 1 t l X l0 ), which is equal to
the numerator of this ratio, since the denominator has been set to 1. In Model (13.18)
the weight associated with Division k is required to equal ω(k), the weight
pre-specified by the decision maker to calculate the system efficiency in Model
(13.17). Since this weight is fixed in Model (3.18), instead of being allowed to vary,
as in Model (13.10), the system efficiency obtained from the former is less than or
equal to that measured from the latter. This result is also reflected from the less
stringent constraint of Model (13.10) than that of Model (13.18).
Applying Model (13.18) to the data contained in Table 13.2 obtains the system
and division efficiencies shown in columns two to four of Table 13.3. As expected,
the system efficiencies in column two are less than those calculated from Model
(13.10) shown in column seven of Table 13.2, as Model (13.18) fixes the weights
ω(1) and ω(2) at 0.5, while Model (13.10) allows them to vary. Moreover, the two
division efficiencies yield an average equal to the system efficiency for every DMU.
For example, the average of the two division efficiencies for DMU A, 0.6666466 and
0.6, is equal to its system efficiency of 0.6333233.
To compare with the results calculated from Model (13.18), we apply Model
(13.17) to the same data, and obtain the results shown in the last three columns of
Table 13.3. Since Models (13.17) and (13.18) have the primal-dual relationship,
their objective values at optimality are the same, and this is reflected from the values
in columns two and five under the headings of E0 and θ, respectively. Although these
two models produce the same system efficiency, their division efficiencies are not
ð1Þ
the same. This is revealed by comparing E0 in column three with θ(1) in column six,
ð2Þ
and E 0 in column four with θ(2) in the last column. One thing special to Model
(13.17) is that when more than one distance parameter is involved it is possible that
some of them may have a value exceeding 1, and a constraint of θ(k) ≤ 1 must be
imposed to avoid obtaining unreasonable results. DMU B is a case of this situation,
where θ(1) has a value of 1.32, which is greater than 1. After adding the constraints of
θ(k) ≤ 1, k = 1, 2, the two division efficiencies change to 1 and 0.96, and the system
efficiency increases to 0.97976. In this example, the non-Archimedean number ε is
set to 0.0001, and the objective value should be adjusted downward by the slacks of
the constraints multiplied by this value, which is why the average of the two division
efficiencies, (θ(1) + θ(2))/2, is slightly higher than the system efficiency for most of
the five DMUs.
This example shows that the division parameter θ(k) is another way of measuring
division efficiencies. Its geometric meaning however is not yet clear, and this
demands further study.
13.3.2.3 Directional Distance Parameter
The input distance parameter shows the excessive inputs that can be reduced while
maintaining the outputs at the current level for an inefficient DMU. In contrast, the
output distance parameter indicates the insufficient outputs that should have been
produced using the current level of inputs. Sometimes it is desired to reduce the input
and increase the output at the same time, and in this case, the directional distance
parameter is a possible choice.
The directional distance function approach is to measure the distance that can be
moved from the current position of a DMU in the input-output space along a certain
direction, which indicates the rate that each input to be reduced and each output to be
increased, to a point on the frontier. For a parallel system with the structure shown in
Fig. 13.4, the directional distance parameter η is measured as:
max: η
n
s:t: λj X ij ≤ X i0 - ηf i , i = mðk - 1Þ þ 1, . . . , mðkÞ , k = 1, . . . , p
j=1
p n
ðkÞ ðkÞ
λj αl X Slj ≤ X Sl0 - ηgl , l = 1, . . . , q
k=1 j=1
n
ðkÞ ðkÞ ðkÞ
λj Y rj ≥ Y r0 þ ηhrðkÞ , r = sðk - 1Þ þ 1, . . . , sðkÞ , k = 1, . . . , p
j=1
ðkÞ
λj ≥ 0, k = 1, . . . , p, j = 1, . . . , n
ð13:19Þ
where (-f, -g, h) is the direction that the DMU being evaluated should move along.
Larger values of the distance parameter η indicate that the corresponding DMU is
less efficient, and an efficient DMU has a value of 0 for η. The direction vector is
usually set to - X0 , - XS0 , Y0 :.
The distance parameter η can be different for the input and output factors, and can
also be different for each division. Which one to use depends on the purposes and
assumptions of the study.
13.3.3 Slacks-based Efficiency Measures
Compared to other approaches, the slacks-based approach is relatively new, with the
term first appearing in 2001 (Tone, 2001), and the first network paper published in
2009 (Tone & Tsutsui, 2009). There are thus relatively few studies on network
systems using the slacks-based approach, with even fewer for parallel systems.
The constraints in the slacks-based model show the slacks between the targets and
the observations, and the objective function is the system efficiency, defined as the
ratio of the average distance parameter of the input factors and that of the output
ðk Þ
factors. In this subsection, the superscript (k) for X i and Y ðrkÞ will be omitted if i runs
for its whole range of 1 to m and r for its whole range of 1 to s. Based on the structure
of the parallel system shown in Fig. 13.4, the model is:
1 msi- q ^sl-
1- þ
mþq i = 1 X i0 l = 1 XS
E 0 = min: l0
1 s sþ
1þ r
r = 1 Y r0
s
n
ðkÞ ðkÞ ðkÞ
s:t: λj X ij þ si- = X i0 , i = mðk - 1Þ þ 1, . . . , mðkÞ , k = 1, . . . , p
j=1
p n
ðkÞ ðkÞ
λj αl X Slj þ ^sl- = X Sl0 , l = 1, . . . , q
k=1 j=1
n
ðkÞ ðkÞ ðkÞ
λj Y rj - sþ
r = Y r0 , r = sðk - 1Þ þ 1, . . . , sðkÞ , k = 1, . . . , p
j=1
λ, s ≥ 0
ð13:20Þ
Since the targets of the shared inputs for each division are not known, we are not
able to measure the efficiency of each division. Consequently, there is no way to
decompose the system efficiency into division efficiencies. Similarly, there is no way
to aggregate the division efficiencies to form a system efficiency, based on which an
efficiency aggregation model is constructed.
If there are no shared inputs, we then have the structure shown in Fig. 13.3, and
Model (13.20) is accordingly simplified to:
1 m si-
1- i=1
m X i0
E 0 = min:
1 s sþ
1þ r
r = 1 Y r0
s
n
ðkÞ ðkÞ ðkÞ
s:t: λj X ij þ si- = X i0 , i = mðk - 1Þ þ 1, . . . , mðkÞ , k = 1, . . . , p
j=1
n
ðkÞ ðkÞ ðkÞ
λj Y rj - sþ
r = Y r0 , r = sðk - 1Þ þ 1, . . . , sðkÞ , k = 1, . . . , p
j=1
λ, s ≥ 0
ð13:21Þ
which is an efficiency decomposition model, in that the objective function is the

system efficiency directly defined from the input and output slacks of the system.
The efficiency of the division based on the slacks in the constraints corresponding to
this division is:
m ðk Þ si-
1- 1
mðkÞ - mðk - 1Þ i = mðk - 1Þ þ1 X ðkÞ
ðk Þ
E0 = i0
, k = 1, . . . , p ð13:22Þ
s ðk Þ sþ
1 þ sðkÞ -1sðk - 1Þ r
r = sðk - 1Þ þ1 Y ðkÞ
r0
There will exist some way to decompose the system efficiency E0 into the
ðk Þ
division efficiencies E 0 , although we will not try to derive this here.
In contrast, we can also construct an efficiency aggregation model by defining the
system efficiency as a weighted average of the division efficiencies. By defining the
weight of a division as the proportion of the average distance parameter of the output
factors of this division in the sum of the average distance parameters of the output
factors of all divisions, that is,
s ðk Þ sþ
1 þ sðkÞ -1sðk - 1Þ r
ωðk Þ = r0
, k = 1, . . . , p
p s ðk Þ sþ
pþ 1
k = 1 s ð k Þ - s ð k - 1Þ
r
r0
then the weighted average of the division efficiencies defined in Eq. (13.22)
becomes:
ðkÞ ðkÞ
1
s
sþ 1
m
si-
1þ ðkÞ ðk -
r
ðkÞ
1- ðkÞ ðk - ðkÞ
p p s -s 1Þ
r = sðk - 1Þ þ1 Y m -m 1Þ
i = mðk - 1Þ þ1 X
ðkÞ
ωðkÞ E 0 =
r0 i0
×
sðkÞ sðkÞ
k =1 k =1 p
1 sþ 1 sþ
pþ ðkÞ - sðk - 1Þ
r
ðkÞ
1 þ ðkÞ ðk - 1Þ r
ðkÞ
k =1 s r =s ðk - 1Þ Y
þ1 r0
s -s r=sðk - 1Þ Y
þ1 r0
mðkÞ
p
1 si-
p- ðkÞ ðk - 1Þ ðkÞ
k =1 m - m i = mðk - 1Þ þ1 X i0
=
sðkÞ
p
1 sþ
pþ r
k =1 s - s
ðkÞ ðk - 1Þ ðkÞ
r = sðk - 1Þ þ1 Y r0
ð13:23Þ
The efficiency aggregation model is thus to minimize this function, subject to the
constraints of Model (13.21), the efficiency decomposition model.
For a special case where all divisions have the same number of input and output
factors, i.e., m(k) - m(k - 1) = m(1) and s(k) - s(k - 1) = s(1) for all Division k, then we
have m(k) - m(k - 1) = m/p and s(k) - s(k - 1) = s/p, where m and s are the total
number of the input and output factors, respectively. In this case the weighted
average of the division efficiencies in Eq. (13.23) becomes:
p mðkÞ si-
p- p
k=1 m i = mðk - 1Þ þ1 X ðkÞ 1- 1 m si-
m i = 1 X i0
=
i0
s sþ
pþ p p sðkÞ sþ
r 1þ 1
s
r
r = 1 Y r0
k=1 s r = sðk - 1Þ þ1 Y ðkÞ
r0
which is just the system efficiency defined in Model (13.21). We have thus obtained
a result that the efficiency decomposition and aggregation models are identical.
This analysis indicates that each division in the efficiency aggregation model has
been assigned an equal weight of 1/p. Consequently, each input factor in a division
with m(k) - m(k - 1) factors has a weight of 1/[p(m(k) - m(k - 1))], and each output
factor will have a weight of 1/[p(s(k) - s(k - 1))]. This implies that the factors of a
division with fewer factors have greater weights than those of a division with more
factors.
The systems discussed in the preceding sections assume all divisions operate
independently, because there are no links between any two divisions. In reality, the
divisions may affect each other. For example, the outpatients of a hospital are
potential inpatients, as they may need to be hospitalized for further treatment after
the preliminary diagnosis as outpatients. Moreover, the inpatients may become
outpatients for periodic check-ups after they are discharged from the hospital. In
this case, we have a linked parallel system. Structurally, this system is the same as
the general series system with reversal links discussed in the preceding chapter. The
models discussed there can thus be modified and used in a straightforward manner.
13.4 Hierarchical Systems 347
13.4 Hierarchical Systems
A hierarchical system can be classified into two types according to whether each
division performs the same function or not, and thus whether it is a multi-component
or multi-function system. In a multi-component system the divisions at all levels
perform the same function. This type of system usually appears with territorial
departmentalization, where the number of divisions at each level need not be the
same for different DMUs. In a multi-function system each division has different
functions. The major characteristic of this type of system is that every level must
have the same number of divisions, and the same division of every DMU has the
same number of subordinate divisions performing distinctive functions. In other
words, there is a one-to-one correspondence of all the divisions of every DMU, and
this is required for them to have the same basis for comparison. The following two
subsections discuss these two cases.
13.4.1 Multi-component Systems
The multi-component hierarchical system is a network system with several levels of

divisions, where all divisions apply the same inputs to produce the same outputs. A
typical example is the power company discussed in Cook and Green (2005), which
owns several power plants, and each power plant has several power units. Specifi-
cally, Power Company j allocates the inputs Xij, i = 1, . . ., m, to the pj subordinate
ðk Þ ðk Þ
plants, and Plant k allocates the inputs X ij it receives from the company to the qj
ðk,dÞ
subordinate power units. Unit d of Plant k consumes the inputs X ij it receives from
ðk,d Þ
its mother plant k to produce the outputs Y rj , r = 1, . . ., s. The total outputs of the
ðk Þ
qj power units subordinate to Plant k are those of this plant, and the total outputs of
pj ðk Þ
the pj plants are those of the company. In symbols, we have X ij = k = 1 X ij and
ðk Þ
ðk Þ qj ðk,dÞ pj ðk Þ
X ij = d = 1 X ij , k = 1,. . ., pj from the input side, and Y rj = k = 1 Y rj and
ðk Þ
ðk Þ qj ðk,d Þ
Y rj = k = 1,. . ., pj from the output side. Figure 13.6 shows the
d = 1 Y rj ,
structure of this hierarchical system.
Suppose there are n companies. The black-box model for measuring the effi-
ciency of a company, indexed as 0, under constant returns to scale, which is the CCR
model, is:
s
r = 1 ur Y r0
E0 = max: m
i = 1 vi X i0
s m ð13:24Þ
s:t: ur Y rj - vi X ij ≤ 0, j = 1, . . . , n
r=1 i=1
ur , vi ≥ ε, 8r, i
a
Company j
Plant 1 Plant 2 Plant pj Level 1
Unit 1-1 ... Unit 1 - q j

(1)
Unit 2-1
... Unit 2 - q ( 2) ... Unit p -1 ... Unit p - q ( p ) Level 2 j
j j j j
b
Company j X ij , i = 1, ..., m
Plant 1 X ij(1) Plant 2 X ij( 2) Plant pj X ij

( pj )
X ij(1,1) X ij
(1, q (j1) ) X ij( 2,1) X ij
( 2, q (j2 ) )
X ij
( p j ,1)
X ij
(p )
( p j ,q j j )
Unit 2-1 ... Unit 2 - q j ...

(1) ( 2)
Unit 1–1 ... Unit 1 - q j
( pj )
Unit pj -1 ... Unit p j - q j
( 2 , q (j2 ) )
Yrj(1,1) (1, q (j1) )
Yrj Yrj( 2,1) Yrj Y rj
( p j ,1) (p )
( p j ,q j j )
Y rj
Yrj(1) Yrj( 2 ) ( pj )
Yrj
Yrj , r = 1, ..., s
Fig. 13.6 A two-level multi-component hierarchical system. (a) Structure, (b) Production system
Obviously, if the operations of the plants and units are not considered, then it is
possible that a company is efficient while many of its plants and units are not. To
eliminate this unreasonable phenomenon, the internal structure of the system must
be taken into account.
Based on the idea of the relational model of Kao (2009a), that the same factor
should have the same multiplier associated with it, no matter which division it
corresponds to, the network model for the hierarchical system of Fig. 13.6 is:
s
r = 1 ur Y r0
E 0 = max: m
i = 1 vi X i0
s:t: Company constraints
s m
ur Y rj - vi X ij ≤ 0, j = 1, . . . , n
r=1 i=1
Plant constraints
s m
ðkÞ ðkÞ
ur Y rj - vi X ij ≤ 0, k = 1, . . . , pj , j = 1, . . . , n
r=1 i=1
Unit constraints
s m
ðk,dÞ ðk,dÞ ðkÞ
ur Y rj - vi X ij ≤ 0, d = 1, . . . , qj , k = 1, . . . , pj , j = 1, . . . , n
r=1 i=1
ur , vi ≥ ε, 8r, i
ð13:25Þ
The objective function is nonlinear, but it can be linearized by assigning a value

of one to the denominator as a constraint, and leaving the numerator as the objective
function. Since the inputs and outputs of the company are the sums of those of its
subordinate plants, the constraint corresponding to the company is the sum of those
s m s pj ðk Þ
corresponding to the plants: r = 1 ur Y rj - i = 1 vi X ij = r = 1 ur k = 1 Y rj -
m pj ðk Þ pj s ðk Þ m ðk Þ
i = 1 vi k = 1 X ij = k=1 r = 1 ur Y rj - i = 1 vi X ij : The company constraints
are thus redundant, and can be deleted. Similarly, since the inputs and outputs of
each Plant k are the sums of those of its subordinate units, the constraint
corresponding to this plant is the sum of those corresponding to its subordinate
ðk Þ ðk Þ
s ðk Þ m ðk Þ s qj ðk,dÞ m qj ðk,dÞ
units: r = 1 ur Y rj - i = 1 vi X ij = r = 1 ur d = 1 Y rj - i = 1 vi d = 1 X ij =
ðk Þ
qj s ðk,d Þ m ðk,d Þ
d=1 r = 1 ur Y rj - i = 1 vi X ij . Thus, the constraints corresponding to the
plants are also redundant and can be omitted. These two levels of redundancy
indicate that only the operations of the divisions at the lowest level need to be
considered in measuring efficiencies. In other words, we have a parallel model
composed of the divisions at the bottom of the hierarchical structure. How the
divisions at the bottom of the structure form the divisions at higher levels is actually
not relevant in measuring the system efficiency, as the corresponding constraints are
redundant.
After an optimal solution (u*, v*) is obtained, the system and division (at all
levels) efficiencies, based on the constraints of Model (13.25), are calculated as:
s
r = 1 ur Y r0
E0 = m
i = 1 vi X i0
s ðkÞ
ðkÞ r = 1 ur Y r0
E0 = , k = 1, . . . , p0 ð13:26Þ
m ðkÞ
i = 1 vi X i0
s ðk,dÞ
ðk,dÞ r = 1 ur Y r0 ðkÞ
E0 = , d = 1, . . . , q0 , k = 1, . . . , p0
m ðk,dÞ
i = 1 vi X i0
Following the efficiency decomposition for parallel systems, the efficiency of a

plant can be expressed as a weighted average of the efficiencies of its subordinate
units and the company efficiency can be expressed as a weighted average of the plant
efficiencies. If we define the weight associated with the subordinate Unit d of Plant
k as the proportion of the aggregate input consumed by this unit in that consumed by
ðk Þ
ðk,dÞ
q ðk,dÞ
Plant k, i.e., ω(k, d )= m
i = 1 vi X i0 / d0= 1 m
i = 1 vi X i0 : Then the weighted average
of the efficiencies of the subordinate power units of Plant k is:
ðkÞ ðkÞ
q0 q0 m ðk,dÞ s ðk,dÞ
ðk,dÞ i = 1 vi X i0 r = 1 ur Y r0
ωðk,dÞ E 0 = ×
d=1 d=1 q0
ðkÞ
ðk,dÞ
m ðk,dÞ
i = 1 vi X i0
m
d=1 i = 1 vi X i0
ðkÞ ðkÞ
q0 s ðk,dÞ q0 s ðk,dÞ
r = 1 ur Y r0 d=1 r = 1 ur Y r0 ð13:27Þ
= ðkÞ
= ðkÞ
d=1 q0 m ðk,dÞ q0 m ðk,dÞ
d=1 i = 1 vi X i0 d=1 i = 1 vi X i0
s ðkÞ
r = 1 ur Y r0 ðkÞ
= = E0
m ðkÞ
i = 1 vi X i0
In other words, the efficiency of a plant (at the first level) is a weighted average of
those of its subordinate units (at the second level).
Similarly, the company efficiency is E0= sr = 1 ur Y r0 = m
i = 1 vi X i0 : If we define
the weight associated with Plant k as the proportion of the aggregate input consumed
ðk Þ
by this plant in that consumed by all plants, i.e., ωðkÞ = m i = 1 vi X i0 /
p0 m ð k Þ
k=1 i = 1 vi X i0 , then the system efficiency can be expressed as:
s s p0 ðkÞ p0 s ðkÞ
r = 1 ur Y r0 r = 1 ur k = 1 Y r0 k=1 r = 1 ur Y r0
E0 = m = ðkÞ
=
m p0 p0 m ðkÞ
i = 1 vi X i0 i = 1 vi k = 1 X i0 k=1 i = 1 vi X i0
ð13:28Þ
p0 m ðkÞ s ðkÞ p0
i = 1 vi X i0 r = 1 ur Y r0 ðkÞ
= × = ωðkÞ E0
k=1
m p0 ðkÞ m ðkÞ k=1
i=1 k = 1 vi X i0 i = 1 vi X i0
A result that the company efficiency is a weighted average of the plant efficien-
cies is obtained.
Combining Eqs. (13.27) and (13.28), the company efficiency can be further
decomposed as:
ðkÞ
p0 p0 q0
ðkÞ ðk,dÞ
E0 = ωðkÞ E0 = ωðkÞ ωðk,dÞ E 0
k =1 k =1 d =1
ðkÞ
p0 m ðkÞ q0 m ðk,dÞ s ðk,dÞ
i=1 vi X i0 i=1 vi X i0 r =1 ur Y r0
= m ×
ðkÞ m ðk,dÞ
k =1 i=1 vi X i0 d =1 q0 m ðk,dÞ
i=1 vi X i0
d =1 vi i=1 X i0
ðkÞ
p0 q0 m ðkÞ m ðk,dÞ s ðk,dÞ
= m ×
ðkÞ m ðk,dÞ
k =1 d =1 i=1 vi X i0 q0 m ðk,dÞ
i=1 vi X i0
d =1 vi i=1 X i0
ðkÞ ðkÞ
p0 q0 m ðkÞ m ðk,dÞ s ðk,dÞ p0 q0
= m × = ^ ðk,dÞ Eðk,dÞ
ω
ðkÞ m ðk,dÞ 0
k =1 d =1 i=1 vi X i0 q0 m ðk,dÞ
i=1 vi X i0 k =1 d =1
d =1 vi i=1 X i0
ð13:29Þ
where
m ðk Þ m ðk,dÞ
i = 1 vi X i0 i = 1 vi X i0
ωðk,dÞ = m × ðk Þ
i = 1 vi X i0 q0 m ðk,d Þ
d = 1 vi i = 1 X i0
with
ðkÞ ðkÞ
p0 q0 p0 q0 m ðkÞ m ðk,dÞ
i = 1 vi X i0 i = 1 vi X i0
^ ðk,dÞ =
ω m × ðkÞ
k=1 d=1 k=1 d=1 i = 1 vi X i0 q0 m ðk,dÞ
d = 1 vi i = 1 X i0
ðkÞ
q0 m ðkÞ m ðk,dÞ m ðkÞ
d = 1ð i = 1 vi X i0 Þð i = 1 vi X i0 Þ
p0 p0
i = 1 vi X i0
= ðkÞ
= m =1
k=1 ð m q0 m ðk,dÞ k=1 i = 1 vi X i0
i = 1 vi X i0 Þ d = 1 vi i = 1 X i0
and ωðk,dÞ ≥ 0: In other words, the company efficiency can be further decomposed
into a weighted average of the efficiencies of the units at the second level.
The above discussion is for hierarchical systems with two levels. The idea can be
extended to systems with three or more levels. It will not be difficult to derive the
conclusion that the system efficiency is a weighted average of the efficiencies of
divisions at the bottom of the structure.
The multi-component hierarchical system is a homogeneous one, in that every
division, no matter at which level it is located, applies the same inputs to produce the
same outputs, and the production frontier is constructed from all the divisions. The
divisions at all levels are thus comparable. This actually follows the idea presented in
Koopmans (1997), which stated that the feasible set in the space of commodity flows
for a firm is the convex hull of the set of vectors resulting from consolidating all
combinations of its production units. Those units at higher levels can be considered
as of greater scale. However, due to possibly multiple solutions, the division
efficiencies calculated from Eq. (13.26) may not be comparable. If two divisions
are to be compared, then one should replace the objective function in Model (13.25)
with the definition of the efficiency of the two divisions separately to calculate their
efficiencies. The efficiencies thus calculated then have a common basis for a fair
comparison.
13.4.2 Multi-function Systems
Different from the multi-component system in which every division performs the
same function by applying the same inputs to produce the same outputs, the multi-
function system is composed of a number of divisions performing different func-
tions, where different inputs are applied to produce different outputs. One example
of this type of system is Chunghwa Post in Taiwan. It has two departments in every
branch, postal and banking services, where the latter is further divided into postal
saving and postal simple life insurance divisions. The branch manager is concerned
with the performance of the two departments, and the head of the banking depart-
ment is concerned with the performance of the two subordinate divisions. The postal
and banking departments are not comparable, and the postal saving and postal
simple life insurance divisions are not comparable, either, and can only be compared
across branches. Since the system efficiency is a function of the division efficiencies,
each branch is seeking the multipliers that will yield the largest efficiency of the
system. In measuring the system efficiency, the division efficiencies will be obtained
at the same time.
The multi-function hierarchical system does not have a standard structure. To get
an idea of how to measure the efficiency of such a system, it is better to look at an
example. Consider the fairly representative structure shown in Fig. 13.7, where the
system is composed of two departments at the first level. These two departments
have three and two divisions, respectively, at the second level, denoted as Div. 1-1,
Div. 1-2, Div. 1-3, Div. 2-1, and Div. 2-2, and Division 2-2 also has two sections at
the third level, denoted as Sec. 2-2-1 and Sec. 2-2-2. For the purpose of easy
referencing, we use superscripts to represent the divisions in concern in the system.
ð1Þ ð1Þ
There are five inputs Xij, i = 1,. . ., 5, supplied to this system, with X 1j , X 2j , and
X 3j allocated to Department 1, and X 3j , X 4j , and X 5j allocated to Department
ð1Þ ð1Þ
2. Based on the subordinate relationship, we have X 1j = X 1j , X 2j = X 2j , X 3j =
X 3j þ X 3j , X 4j = X 4j , and X 5j = X 5j , where input X3j is shared between these two
ð1Þ ð2Þ
departments. In vector forms, we have X j = X j þ X j : Each division distributes
the inputs allocated to it from its mother division to its subordinate divisions for
further distribution, until the terminal divisions, where the inputs they receive are
consumed to produce outputs. For those non-terminal divisions which do not
a
Company
Dept. 1 Dept. 2 Level 1
Div. 1-1 Div. 1-2 Div. 1-3 Div. 2-1 Div. 2-2 Level 2
Sec. 2-2-1 Sec. 2-2-2 Level 3
b X ij , i = 1, ..., 5
DMU j
X 1(1j ) , X 2(1j) , X 3(1j) X 3( 2j ) , X 4( 2j ) , X 5( 2j )
Dept. 1 Dept. 2
X 1(1j,1) , X 2(1j,1) X 2(1j,2) , X 3(1j,2) X 1(1j,3) , X 3(1j,3) X 3( 2j ,1) , X 4( 2j,1) X 3( 2j ,2) , X 5( 2j ,2)
Div. 1-1 Div. 1-2 Div. 1-3 Div. 2-1

X 3( 2j ,2,1) X 5( 2j ,2,2)
Y1(j1,1) Y2(1j ,2) Y3(1j ,3) Y4( 2j ,1) Sec. 2-2-1 Sec. 2-2-2
Y5(j2,2,2) Y6( j2,2,2)
Y5(j2,2) , Y6( j2,2)
Y1(j1) , Y2(1j ) , Y3(1j ) Y4( j2 ) , Y5(j2 ) , Y6( j2)
Yrj , r = 1, ..., 6
Fig. 13.7 A three-level multi-function hierarchical system. (a) Structure, (b) Production system
produce outputs themselves, their outputs are those produced by their subordinate
divisions. For example, the outputs of Division 2-2 are those produced by Sections
2-2-1 and 2-2-2, and those of Department 2 are produced by Divisions 2-1 and 2-2.
ð2,2Þ ð2,2,1Þ ð2,2Þ ð2,2,2Þ ð2Þ ð2,1Þ ð2Þ
In symbols, we have Y 5j = Y 5j , Y 6j = Y 6j , and Y 4j = Y 4j , Y 5j =
ð2,2Þ ð2Þ ð2,2Þ ð2,2Þ
Y 5j , Y 6j = Y 6j : This can also be expressed more concisely as Y j =
ð2,2,1Þ ð2,2,2Þ ð2Þ ð2,1Þ ð2,2Þ
Yj þ Yj and Yj þ = Yj Y j :.
Suppose there are n DMUs of the same structure to be compared when measuring
their relative efficiency. The hierarchical model for measuring the efficiency of this
system is used to maximize the efficiency of the system, taking the operations of the
divisions into account. The associated model is:
6
r = 1 ur Y r0
E 0 = max: 5
i = 1 vi X i0
s:t: System constraints :
6 5
ur Y rj - vi X ij ≤ 0, j = 1, . . . , n
r=1 i=1
Department constraints :
3 3 6 5
ur Y rj - vi X ij ≤ 0, ur Y rj - vi X ij ≤ 0, j = 1, . . . , n
r=1 i=1 r=4 i=3
ð1,1Þ ð1,1Þ ð1,1Þ ð1,2Þ ð1,2Þ ð1,2Þ
u1 Y 1j - v1 X 1j þ v2 X 2j ≤ 0, u2 Y 2j - v2 X 2j þ v3 X 3j ≤ 0,
ð1,3Þ ð1,3Þ ð1,3Þ ð2,1Þ ð2,1Þ ð2,1Þ

u3 Y 3j - v1 X 1j þ v3 X 3j ≤ 0, u4 Y 4j - v3 X 3j þ v4 X 4j ≤ 0,
ð2,2Þ ð2,2Þ ð2,2Þ ð2,2Þ

u5 Y 5j þ u6 Y 6j - v3 X 3j þ v5 X 5j ≤ 0, j = 1, . . . , n
Section constraints :
ð2,2,1Þ ð2,2,1Þ ð2,2,2Þ ð2,2,2Þ
u5 Y 5j - v3 X 3j ≤ 0, u6 Y 6j - v5 X 5j ≤ 0, j = 1, . . . , n
ur , vi ≥ ε, r = 1, . . . , 6, i = 1, . . . , 5
ð13:30Þ
After an optimal solution (u, v) is obtained, different efficiencies of this system
are calculated as:
System efficiency:
6
r = 1 ur Y r0
E0 = ð13:31Þ
5
i = 1 vi X i0
Department efficiencies:
3 ð1Þ 6 ð2Þ
ð1Þ r = 1 ur Y r0 ð2Þ r = 4 ur Y r0
E0 = , E0 =
3 ð1Þ 5 ð2Þ
i = 1 vi X i0 i = 3 vi X i0
Division efficiencies:
ð1,1Þ ð1,2Þ ð1,3Þ

ð1,1Þ u1 Y 10 ð1,2Þ u2 Y 20 ð1,3Þ u3 Y 30
E0 = ð1,1Þ ð1,1Þ
, E0 = ð1,2Þ ð1,2Þ
, E0 = ð1,3Þ ð1,3Þ
v1 X 10 þ v2 X 20 v2 X 20 þ v3 X 30 v1 X 10 þ v3 X 30
ð2,1Þ ð2,2Þ ð2,2Þ
ð2,1Þ u4 Y 40 ð2,2Þ u5 Y 50 þ u6 Y 60
E0 = ð2,1Þ ð2,1Þ
, E0 = ð2,2Þ ð2,2Þ
v3 X 30 þ v4 X 40 v3 X 30 þ v5 X 50
Section efficiencies:
ð2,2,1Þ ð2,2,2Þ
ð2,2,1Þ u5 Y 50 ð2,2,2Þ u6 Y 60
E0 = ð2,2,1Þ
, E0 = ð2,2,2Þ
v3 X 30 v5 X 50
Note that the conventional CCR black-box model only has the system constraints.
The system efficiency calculated from the Hierarchical Model (13.30) is obviously
less than or equal to that calculated from the black-box model.
Similar to that discussed in multi-component systems, only the constraints
corresponding to the terminal divisions are required, as the other constraints whose
corresponding divisions have subordinate divisions are redundant, and thus can be
deleted. For example, Division 2-2 has two subordinate divisions, Sec. 2-2-1 and
ð2 - 2Þ ð2 - 2 - 1Þ ð2 - 2 - 2Þ ð 2 - 2Þ
Sec. 2-2-2, and we thus have X j = Xj þ Xj and Y j =
ð2 - 2 - 1Þ ð2 - 2 - 2Þ ð2 - 2Þ ð2 - 2Þ
Yj þ Yj : This shows that uY j - vX j =
ð2 - 2 - 1Þ ð2 - 2 - 1Þ ð2 - 2 - 2Þ ð2 - 2 - 2Þ
uY j - vX j þ uY j - vX j , which indicates that
the constraint corresponding to Division 2-2 is redundant, and can be deleted.
Taking this into consideration, it turns out that only six constraints corresponding
to the terminal divisions of Div. 1-1, Div. 1-2, Div. 1-3, Div. 2-1, Sec. 2-2-1, and
Sec. 2-2-2 are needed in Model (13.30). When a division has subordinate divisions,
it can be considered as a system with its subordinate divisions arranged in a parallel
structure. In this case, the efficiency of the division will be a weighted average of
those of the subordinate divisions, where the weight of the subordinate division is the
proportion of the aggregate input consumed by this subordinate division in that
consumed by all subordinate divisions. For example, Division 2-2 has two subordi-
nate divisions of Sec. 2-2-1 and Sec. 2-2-2. The efficiency of Division 2-2 is then a
weighted average of those of Sec. 2-2-1 and Sec. 2-2-2, as derived below:
ð2,2,1Þ ð2,2,1Þ
ð2,2,1Þ ð2,2,2Þ v3 X 30 u5 Y 50
ωð2,2,1Þ E0 þ ωð2,2,2Þ E0 = ð2,2,1Þ ð2,2,2Þ
× ð2,2,1Þ
þ
v3 X 30 þ v5 X 50 v3 X 30
ð2,2,2Þ ð2,2,2Þ ð2,2,1Þ ð2,2,2Þ
v5 X 50 u6 Y 60 u5 Y 50 þ u6 Y 60
ð2,2,1Þ ð2,2,2Þ
× ð2,2,2Þ
= ð2,2,1Þ ð2,2,2Þ
v3 X 30 þ v5 X 50 v5 X 50 v3 X 30 þ v5 X 50
ð2,2Þ ð2,2Þ
u5 Y 50 þ u6 Y 60 ð2,2Þ
= ð2,2Þ ð2,2Þ
= E0
v3 X 30 þ v5 X 50
where ω(2, 2, 1) + ω(2, 2, 2) = 1. By the same token, the efficiencies of Departments

ð1Þ ð1,1Þ
1 and 2 and the system as a whole can be expressed as E 0 = ωð1,1Þ E0 þ
ð1,2Þ ð1,3Þ ð2Þ ð2,1Þ ð2,2Þ ð1Þ
ωð1,2Þ E 0 þ ωð1,3Þ E0 , E 0 = ωð2,1Þ E 0 þ ωð2,2Þ E 0 , and E 0 = ωð1Þ E 0 þ
ð Þ
ωð2Þ E 0 , respectively. Based on these decompositions, the system efficiency E0
2
can be decomposed as a weighted average of the efficiencies of the divisions at a

lower level, and this process can be continued, level by level, down to the very
ð1,1Þ ð1,2Þ ð1,3Þ ð2,1Þ ð2,2,1Þ ð2,2,2Þ
bottom of the six divisions, E 0 , E 0 , E 0 , E 0 , E 0 , and E 0 :

E0 =ωð1Þ E 0 þωð2Þ E0 =ωð1Þ ωð1,1Þ E 0 þωð1,2Þ E 0 þωð1,3Þ E0 þ
ð2,1Þ ð2,2Þ ð1,1Þ ð1,2Þ ð1,3Þ

ωð2Þ ωð2,1Þ E 0 þωð2,2Þ E0 =ωð1Þ ωð1,1Þ E 0 þωð1,2Þ E 0 þωð1,3Þ E0 þ
ð2,1Þ ð2,2,1Þ ð2,2,2Þ

ωð2Þ ωð2,1Þ E 0 þωð2,2Þ ωð2,2,1Þ E 0 þωð2,2,2Þ E 0
ð1,1Þ ð1,2Þ ð1,3Þ ð2,1Þ ð2,2,1Þ ð2,2,2Þ
^ ð1,1Þ E 0
=ω ^ ð1,2Þ E 0
þω ^ ð1,3Þ E0
þω ^ ð2,1Þ E 0
þω ^ ð2,2,1Þ E0
þω ^ ð2,2,2Þ E0
þω
where
ωð1,1Þ = ωð1Þ ωð1,1Þ , ωð1,2Þ = ωð1Þ ωð1,2Þ , ωð1,3Þ = ωð1Þ ωð1,3Þ , ωð2,1Þ = ωð2Þ ωð2,1Þ ,
ωð2,2,1Þ = ω(2)ω(2, 2)ω(2, 2, 1), and ωð2,2,2Þ = ωð2Þ ωð2,2Þ ωð2,2,2Þ , and it is noted that
ωð1,1Þ þ ωð1,2Þ þ ωð1,3Þ þ ωð2,1Þ þ ωð2,2,1Þ þ ωð2,2,2Þ =
ð Þ ð Þ ð1,2Þ ð1,2Þ
v1 X 10 þ v2 X 20 = m
v2 X 20 þ v3 X 30 Þ= m
1,1 1,1
i = 1 vi X i0 + i = 1 vi X i0 +
ð1,3Þ ð1,3Þ ð2,1Þ ð2,1Þ
v1 X 10 þ v3 X 30 Þ= m
i = 1 vi X i0 + v3 X 30 þ v4 X 40 = m
i = 1 vi X i0 +
ð2,2,1Þ ð2,2,2Þ
v3 X 30 = m
i = 1 vi X i0 + v5 X 50 = m
i = 1 vi X i0 = 1. This indicates that the system
efficiency can actually be decomposed into a weighted average of the efficiencies of the
terminal divisions, and the weight associated with a terminal division is the aggregate
input consumed by this terminal division in that consumed by all terminal divisions. In
other words, a multi-function hierarchical system is equivalent to a parallel system
composed of terminal divisions.
Measuring the efficiency of parallel systems can be traced back to the work of Färe
and Primont (1984), which studied the efficiency measurement of multi-plant firms,
although the idea of parallel systems was not clear at that time. Kao (1998) followed
the same idea to measure the efficiency of the 34 working circles subordinated to the
13 forest districts in Taiwan.
Most of the studies of parallel systems are concerned with multi-function systems
with shared input and/or output factors. In measuring the performance of the
chemistry and physics departments in UK universities, Beasley (1995) developed
a model to aggregate the efficiency of teaching and research. While the structure of
the system was not expressed as parallel, the model was a parallel one. Yang et al.
(2000) constructed a model to find that the system efficiency is equal to the
maximum of the division efficiencies measured independently from the CCR
model. Cook and Hababou (2001) measured the performance of 20 branches of a
Canadian bank, with sales and services as the two major functions. The model was
able to derive an optimal split of the shared resources that maximized the system
efficiency. Cook et al. (2000) studied the same problem, although the same resource
used by different divisions was assumed to have different values, and thus different
multipliers were attached. Cook and Green (2004) followed the same idea to study
the performance of 10 steel plants, where each plant had four components for
production.
Jahanshahloo et al. (2004a) evaluated the performance of 39 branches of a
commercial bank in Iran, in which the functions were classified as deposits, sales,
and services, and both shared inputs and outputs were involved. By requiring the
same factor to have the same multiplier, the system efficiency was found to be a
weighted average of the division efficiencies. Jahanshahloo et al. (2004b) extended
this model to incorporate non-discretionary inputs. Amirteimoori and Kordrostami
(2005) built a model to study a case with imprecise data, i.e., the data were of
bounded value and ordinal ranks. Amirteimoori and Nashtaei (2006) also classified
the functions of a bank into deposits, sales, and services, to measure the efficiency of
28 branches of an Iranian bank. Bian et al. (2015) measured 18 railway firms in
China, with transporting passengers and freight as the two major functions.
Kao (2009b) proposed a relational model to measure the system and division
efficiencies in one model for parallel systems, with the property that the system
efficiency is a weighted average of the component division efficiencies. Kao (2012)
applied the same model to measure the teaching and research efficiencies of
52 chemistry departments in the UK (using the data compiled by Beasley (1995)).
Kao and Lin (2011) measured efficiency when the factors were qualitative, while
Kao and Lin (2012) measured it when the observations were fuzzy numbers. Lozano
(2014) developed a similar model for handling fuzzy data based on the idea
presented in Kao and Lin (2012). Bi et al. (2011) followed the main idea of Kao
(2009b), with some modifications, to study problems related to resource allocation
and target setting. A set of common weights was chosen such that the worst behaved
unit was improved as much as possible. Lei et al. (2015) assessed the achievements
of 25 countries in the Olympic Games by considering the Summer and Winter
Olympics as two parallel games, with GDP per capita and population as the shared
inputs and the three types of medal as the shared outputs. Kao and Liu (2022) studied
a group decision-making problem where each expert was considered as a division in
a parallel system. Their model shows that aggregating the opinions of all experts to
calculate efficiencies obtains the same result as aggregating the efficiencies calcu-
lated from the opinions of individual exerts. The final efficiencies are thus reliable.
Rogge and Jaeger (2012) analyzed cost efficiency in the treatment of solid waste
in 293 municipalities in Flanders, Belgium, using a ratio-form system efficiency
model. There were six types of solid waste, residential, other municipal, packaging,
other EPR (Extended Producer Responsibility), green, and bulky, with a shared
input of handling costs. The results showed that the cost efficiency of the system is a
weighted average of those of the divisions. Da Cruz et al. (2013) applied the same
model to measure the efficiency of the drinking water and wastewater services of
45 water utilities in Portugal with shared resources.
Färe et al. (1997) is one of the earliest works on parallel systems, and their model
aims to maximize the output system distance parameter. In their study of 57 Southern
Illinois grain farms, the land was shared for crops of corn, soybeans, wheat, and
double crop soybeans. The solution not only showed the efficiency of each farm, but
also the best allocation of land for each crop. Bi et al. (2012) classified the production
activities of a business into core and non-core ones, operating in a parallel manner.
The inputs were shared, and the outputs were the contributions of both divisions. The
conventional input system distance function was used to assess 20 convenience
stores by applying a parametric bootstrap method. Yu (2008) studied the perfor-
mance of multi-mode transit firms, where the bus services were separated into
highway and urban ones. A division distance measurement model was then used
to measure the performance of 60 bus companies in Taiwan. Tsai and Mar Molinero
(2002) investigated the performance of 27 acute hospital trusts in the UK under five
categories, medical, surgical, maternity, psychiatric, and others, using a division
distance measurement model. Diez-Ticio and Mancebon (2002) applied the ideas
presented in Beasley (1995), Mar Molinero (1996), and Mar Molinero and Tsai
(1997) to analyze the efficiency of the Spanish Police Service. The services were
separated into “crime control” and “protection and maintenance of order,” with the
number of policemen as a shared input, and observations from 47 provincial capitals
were collected for measurement. Chao et al. (2010) separated the functions of
financial holding companies into banking, insurance, security, and others, and
measured the performance of 12 financial holding companies in Taiwan. Shared
inputs were also incorporated to find the most favorable share for each division in the
measurement.
Directional distance measures have also been applied to measure efficiency for
parallel systems when undesirable outputs appear. Yu and Fan (2006) studied the
performance of multi-mode transit firms, where the bus services were separated into
highway and urban ones. This model was applied to measure the performance of
24 bus companies in Taiwan, with the number of accidents as an undesirable output.
References 359
Grosskopf et al. (2015) investigated budget allocation between the central adminis-
tration and the constituent schools of 70 school districts in the Dallas, Texas, area,
where the schools were separated into primary, middle, mixed, and high.
Slacks-based approaches for measuring the efficiency of parallel systems have
received relatively little attention. Lozano (2015) proposed a model to measure the
efficiency of parallel systems with joint inputs which are used in two processes,
production and pollution-generating, and the efficiencies of 92 coal-fired power
plants were measured for illustration.
Vaz et al. (2010) assessed the performance of 78 retail stores in Portugal with five
sections of groceries, perishables, light bazaar, heavy bazaar, and textiles, using a
value-based model with the aim of maximizing overall sales. The floor area was a
shared input allowing for reallocation.
Hierarchical systems have received relatively little attention, as disclosed in the
review of Kao (2014). Of the limited works that do consider this structure, Cook and
Green (2005) discussed a multi-component hierarchical system composed of several
layers based on the idea presented in Cook et al. (1998), and developed a CCR-type
model to measure the efficiency of 10 Canadian power plants with 40 subordinate
power units, and each power plant does not necessarily have the same number of
power units. Castelli et al. (2004) proposed a model to measure the performance of a
multi-function hierarchical system with one and two levels, where the units at a
higher level can have common subordinate units at a lower level. Since one-level
systems have the same structure as parallel ones, Kao’s (2009b) model for measuring
the efficiency of the latter can also be used with the former. With regard to two-level
systems, Castelli et al. (2004) proposed two models to measure DMU efficiency, one
that carries out weight balancing, which requires the input weight of each unit at a
lower level to be equal to the sum of the output weights of the units at a higher level
flowing into this unit, and one that carries out flow balancing, which requires the
flow entering each unit at a lower level to be equal to the sum of the flows from the
units at a higher level. However, the models of Castelli et al. (2004) only measure the
efficiency of the whole DMU. Kao (2015) discovered that the hierarchical system
can be represented by a parallel system composed of the divisions at the bottom of
the hierarchical structure. Carayannis et al. (2015) analyzed the innovation efficiency
of 23 countries with a multi-level multi-stage approach, where the innovation
efficiency is separated into knowledge production and knowledge commercializa-
tion processes at the national and regional levels.
References
Amirteimoori, A., & Kordrostami, S. (2005). Multi-component efficiency measurement with

imprecise data. Applied Mathematics and Computation, 162, 1265–1277.
Amirteimoori, A., & Nashtaei, R. A. (2006). The role of time in multi-component efficiency
analysis: An application. Applied Mathematics and Computation, 177, 11–17.
Beasley, J. (1995). Determining teaching and research efficiencies. Journal of the Operational
Bi, G. B., Ding, J. J., & Luo, Y. (2011). Resource allocation and target setting for parallel
production system based on DEA. Applied Mathematical Modelling, 35, 4270–4280.
Bi, G. B., Feng, C. P., Ding, J. J., & Khan, M. R. (2012). Estimating relative efficiency of DMU:
Pareto principle and Monte Carlo oriented DEA approach. INFOR, 50, 44–57.
Bian, Y., Hu, M., & Xu, H. (2015). Measuring efficiencies of parallel systems with shared inputs/
outputs using data envelopment analysis. Kybernetes, 44, 336–352.
Carayannis, E. G., Goletsis, Y., & Grigoroudis, E. (2015). Multi-level multi-stage efficiency
measurement: The case of innovation systems. Operational Research, 15, 253–274.
Castelli, L., Pesenti, R., & Ukovich, W. (2004). DEA-like models for the efficiency evaluation of
hierarchically structured units. European Journal of Operational Research, 154, 465–476.
Chao, C. M., Yu, M. M., & Chen, M. C. (2010). Measuring the performance of financial holding
companies. The Service Industries Journal, 30, 811–829.
Cook, W. D., Chai, D., Doyle, J., & Green, R. (1998). Hierarchies and groups in DEA. Journal of
Cook, W. D., & Green, R. H. (2004). Multicomponent efficiency measurement and core business
identification in multiplant firms: A DEA model. European Journal of Operational Research,
157, 540–551.
Cook, W. D., & Green, R. H. (2005). Evaluating power plant efficiency: A hierarchical model.
Computers & Operations Research, 32, 813–823.
Cook, W. D., & Hababou, M. (2001). Sales performance measurement in bank branches. Omega,
29, 299–307.
Cook, W. D., Hababou, M., & Tuenter, H. J. H. (2000). Multicomponent efficiency measurement
and shared inputs in data envelopment analysis: An application to sales and service performance
in bank branches. Journal of Productivity Analysis, 14, 209–224.
Da Cruz, N. F., Carvalho, P., & Marques, R. C. (2013). Disentangling the cost efficiency of jointly
provided water and wastewater services. Utilities Policy, 24, 70–77.
Diez-Ticio, A., & Mancebon, M. J. (2002). The efficiency of the Spanish police service: An
application of the multiactivity DEA model. Applied Economics, 34, 351–362.
Färe, R., Grabowski, R., Grosskopf, S., & Kraft, S. (1997). Efficiency of a fixed but allocatable
input: A non-parametric approach. Economics Letters, 56, 187–193.
Färe, R., & Primont, D. (1984). Efficiency measures for multiplant firms. Operations Research
Letters, 3, 257–260.
Grosskopf, S., Hayes, K., Taylor, L. L., & Weber, W. (2015). Centralized or decentralized control
of school resources? Journal of Productivity Analysis, 43, 139–150.
Jahanshahloo, G. R., Amirteimoori, A. R., & Kordrostami, S. (2004a). Multi-component perfor-
mance, progress and regress measurement and shared inputs and outputs in DEA for panel data:
An application in commercial bank branches. Applied Mathematics and Computation, 151,
1–16.
Jahanshahloo, G. R., Amirteimoori, A. R., & Kordrostami, S. (2004b). Measuring the multi-
component efficiency with shared inputs and outputs in data envelopment analysis. Applied
Mathematics and Computation, 155, 283–293.
Kao, C. (1998). Measuring the efficiency of forest districts with multiple working circles. Journal of
Kao, C. (2009a). Efficiency decomposition in network data envelopment analysis: A relational
Kao, C. (2009b). Efficiency measurement for parallel production systems. European Journal of
Kao, C. (2012). Efficiency decomposition for parallel production systems. Journal of the Opera-
tional Research Society, 63, 64–71.
Kao, C. (2014). Network data envelopment analysis: A review. European Journal of Operational
Kao, C. (2015). Efficiency measurement for hierarchical network systems. Omega, 51, 121–127.
References 361
Kao, C., & Lin, P. H. (2011). Qualitative factors in data envelopment analysis: A fuzzy number
Kao, C., & Lin, P. H. (2012). Efficiency of parallel production systems with fuzzy data. Fuzzy Sets
and Systems, 198, 83–98.
Kao, C., & Liu, S. T. (2022). Group decision making in data envelopment analysis: A robot
selection application. European Journal of Operational Research, 297, 592–599.
Koopmans, T. C. (1997). Examples of production relations based on microdata. In G. C. Harcourt
(Ed.), The microeconomic foundation of macroeconomics (pp. 144–171). Macmillan.
Lei, X., Li, Y., & Xie, Q. (2015). Measuring Olympic achievements based on a parallel DEA
approach. Annals of Operations Research, 226, 379–396.
Lozano, S. (2014). Computing fuzzy process efficiency in parallel systems. Fuzzy Optimization and
Decision Making, 13, 73–89.
Lozano, S. (2015). A joint-inputs network DEA approach to production and pollution-generating
technologies. Expert Systems with Applications, 42, 7960–7968.
Mar Molinero, C. (1996). On the joint determination of efficiencies in a data envelopment analysis
context. Journal of the Operational Research Society, 47, 1273–1279.
Mar Molinero, C., & Tsai, P. (1997). Some mathematical properties of a DEA model for the joint
determination of efficiencies. Journal of the Operational Research Society, 48, 51–56.
Rogge, N., & Jaeger, S. (2012). Evaluating the efficiency of municipalities in collecting and
processing municipal solid waste: A shared input DEA-model. Waste Management, 32,
1968–1978.
Tone, K., & Tsutsui, M. (2009). Network DEA: A slacks-based approach. European Journal of
Tsai, P. F., & Mar Molinero, C. (2002). A variable returns to scale data envelopment analysis model
for the joint determination of efficiencies with an example of the UK health service. European
Vaz, C. B., Camanho, A. S., & Guimarães, R. C. (2010). The assessment of retailing efficiency
using network data envelopment analysis. Annals of Operations Research, 173, 5–24.
Yang, Y., Ma, B., & Koike, M. (2000). Efficiency-measuring DEA model for production system
with k independent subsystems. Journal of the Operations Research Society of Japan, 43,
343–354.
Yu, M. M. (2008). Measuring the efficiency and return to scale status of multi-mode bus transit -
evidence from Taiwan’s bus system. Applied Economics Letters, 15, 647–653.
Yu, M. M., & Fan, C. K. (2006). Measuring the cost effectiveness of multimode bus transit in the
presence of accident risks. Transportation Planning and Technology, 129, 383–407.
Chapter 14
Mixed Systems
Theoretically, network systems can have numerous structures. In practice, however,

only a limited number of them exist, and most of them can be classified into series
and parallel, with only a few left unclassified. The unclassified systems are basically
mixtures of the series and parallel ones, which do not explicitly show the relationship
between the system and division efficiencies. Once the relationship between the
system and division efficiencies is explored, the divisions that have the greatest
impact on the performance of the system can be identified. Improvements to these
divisions will increase the efficiency of the system the most.
Only a few mixed systems are reported in the literature; the models used to
measure the efficiencies of these fall into the broad categories of independent, ratio-
form, distance function, and slacks-based. It is interesting to note that, excluding
those that have been classified, no more than 20 specific structures appear in the
literature, and we will use ten of these structures to explain the models proposed for
measuring the efficiencies of the system and divisions. In addition, we will also
construct the relational model, introduced by Kao (2009), to decompose the effi-
ciency of the system into those of the divisions, so that the divisions that have the
greatest impact on the performance of the system can be identified. The discussion is
separated into three sections according to the number of divisions in the structure,
including four structures of three divisions in Sect. 14.1, four structures of four
divisions in Sect. 14.2, and two structures of five divisions in Sect. 14.3. Finally, in
the Supplementary Literature Sect. 14.4, some applications of the mixed system that
have appeared in the literature are briefly described.
https://doi.org/10.1007/978-3-031-27593-7_14
364 14 Mixed Systems
14.1 Three-Division Structures
The minimum number of divisions in a structure that is neither series nor parallel is
three. In this section we introduce four structures composed of three divisions that
have appeared in the literature.
14.1.1 The R&D Value Chain Example
In an attempt to grasp some idea of whether production and R&D affect the market
value of a firm, Wang et al. (2013) proposed an R&D value chain fraamework to
explore the relationship among productivity, R&D, and firm market values using a
sample of 65 high-technology firms in Taiwan. Eight productivity performance
indicators were considered, including the number of employees (X1), the assets
(X2), the number of researchers (X3), R&D expenditures (X4), sales volumes (Z1),
number of patents (Z2), market values (Y1), and returns on investment (Y2). The value
chain has two stages, which measures the profitability and marketability efficiencies.
Different from the conventional two-stage studies, the first stage is composed of two
divisions: basic operation (Division 1) and R&D effects (Division 2), operating in
parallel. The second stage, as usual, has one division, market (Division 3). The basic
operation division employs employees (X1) and assets (X2) to produce sales (Z1), the
R&D division applies researchers (X3) and R&D expenditures (X4) to produce
patents (Z2), and the market division uses sales (Z1) and patents (Z2) to generate
market values (Y1) and returns (Y2). Figure 14.1 depicts the structure of this system.
The following distance parameter mdel under variable returns to scale was proposed
to measure the efficiency of the three divisions:
Fig. 14.1 The structure of

the R&D value chain X 1, X 2 Z1
problem 1
Y1 ,Y2
3
X 3, X 4 Z2
2
14.1 Three-Division Structures 365
min: w1 θ1 þ w2 θ2 - w3 φ
Division 1
65
ð1Þ
s:t: λj X ij ≤ θ1 X i0 , i = 1, 2
j=1
65
ð1Þ
λj Z 1j ≥ ~z10
j=1
65
ð1Þ
λj = 1, θ1 ≤ 1
j=1
Division 2
65
ð2Þ
λj X ij ≤ θ2 X i0 , i = 3, 4
j=1
65
ð2Þ
λj Z 2j ≥ ~z20 ð14:1Þ
j=1
65
ð2Þ
λj = 1, θ2 ≤ 1
j=1
Division 3
65
ð3Þ
λj Z gj ≤ ~zg0 g = 1, 2
j=1
65
ð3Þ
λj Y rj ≥ φY r0 , r = 1, 2
j=1
65
ð3Þ
λj = 1, φ≥1
j=1
ðkÞ
λj , ~z10 , ~z20 ≥ 0, k = 1, 2, 3; j = 1, . . . , 65
where w1, w2, and w3 are pre-determined weights that reflects the importance of the
associated division. At optimality, θ1 , θ2 , and 1/ϕ are the efficiencies of the three
divisions. Basically, this model is of the independent form (referring to Model (9.7)
in Chap. 9). It is expected that the properties possessed by the independent model
discussed in Chap. 9 also apply here.
Model (14.1) is not able to measure the efficiency of the system. The relationship
between the system and division efficiencies is not clear, either. A closer examina-
tion of the structure shown in Fig. 14.1 reveals that the system is actually a basic two-
stage system where the first stage is composed of two divisions connected in parallel.
A relational model under constant returns to scale can thus be constructed to express
the system efficiency as the product of the two stage efficiencies and the efficiency of
the first stage is a weighted average of the efficiencies of the two component
divisions. The relational model for this system is:
E 0 = max: u1 Y 10 þ u2 Y 20
4
s:t: vi X i0 = 1
i=1
w1 Z 1j - ðv1 X 1j þ v2 X 2j Þ ≤ 0, j = 1, . . . , 65 ð14:2Þ
w2 Z 2j - ðv3 X 3j þ v4 X 4j Þ ≤ 0, j = 1, . . . , 65
u1 Y 1j þ u2 Y 2j - ðw1 Z 1j þ w2 Z 2j Þ ≤ 0, j = 1, . . . , 65
u1 , u2 , vi , w1 , w2 ≥ ε, i = 1, 2, 3, 4
After a set of optimal solutions is obtained, the system and division efficiencies
are calculated as:
u1 Y 10 þ u2 Y 20
E0 = 4
i = 1 vi X i0
ð1Þ w1 Z 10
E0 =
v1 X 10þ v2 X 20
ð2Þ w2 Z 20
E0 =
v3 X 30þ v4 X 40
ð3Þ u1 Y 10 þ u2 Y 20
E0 =
w1 Z 10 þ w2 Z 20
The system efficiency can be decomposed as:
u1 Y 10 þ u2 Y 20 w1 Z 10 þ w2 Z 20 u1 Y 10 þ u2 Y 20

E0 = = ×
4 4 w1 Z 10 þ w2 Z 20
vi X i0 vi X i0
i=1 i=1
v1 X 10 þ v2 X 20 w1 Z 10 v X 30 þ v4 X 40 w Z 20

= × þ 3 4 × 2
4 v1 X 10
þ v2 X 20 v3 X 30 þ v4 X 40
vi X i0 vi X i0
i=1 i=1
u1 Y 10 þ u2 Y 20 ð1Þ ð2Þ ð3Þ
× = ωð1Þ E 0 þ ωð2Þ E 0 × E0
w1 Z 10 þ w2 Z 20
where ω1 = v1 X 10 þ v2 X 20 = 4i = 1 vi X i0 and ω2 = 1 - ω1. If the technology is of

variable returns to scale, then the relationship between the system and three division
efficiencies is not clear, although all efficiencies can be measured.
14.1.2 The Railway Operation Example
In the previous example the two divisions in Stage 1 do not have exogenous outputs
and the division in Stage 2 does not have exogenous inputs, which makes efficiency
decomposition relatively easy. In this subsection we introduce a structure which is a
little more complicated, and a transformation of the structure is needed to make
efficiency decomposition possible.
Yu and Lin (2008) investigated the efficiency of railway services of 20 European
railway companies, where the operation of a company can be separated into two
stages: production and consumption. In the production stage passengers and freight
services are being provided, and the services are consumed by customers. The
characteristic of railway services is that the unused services are vanished and wasted.
The employees (X S1 ) are the workforce of a railway company. They share the works
of passenger production (Division 1), freight production (Division 2), and consump-
tion of services (Division 3). The railway lines (X S2) are used for both passenger and
freight transportation. The numbers of passenger cars (X1) and freight cars (X2) are
used for carrying passengers and freight, respectively. The outputs of passenger
production and freight production are passenger train-km (Z1) and freight train-km
(Z2), respectively, and these two outputs are used by the consumption division to
produce passenger-km (Y1) and ton-km (Y2). The structure of this system is shown in
Fig. 14.2(a).
Fig. 14.2 The basic three a

division structure with X 1S
exogenous inputs. (a)
Original (b) Series-parallel
conversion X1 Z1
1
X 1S , X 2S
Y1
3
X 1S , X 2S Z2 Y2
2
X2
X1
Z1
P1 1
D1
D3 D3 X 1S Y1
X 2S X 1S 4 3
Y2
D2
P2 Z2
X2 2
Let α1, α2, and α3, with α1 + α2 + α3 = 1, be the proportion of the employees
devoted to the works associated with passenger production, freight production, and
services, respectively, and μ1 and μ2, with μ1 + μ2 = 1, be the proportion of the
railway lines dedicated to passenger production and freight production, respectively.
The proportions of αi, i = 1, 2, 3 are confined to the range of [αi , αi ] and that of μi,
i = 1, 2 are confined to the range of [μi , μi ]. Yu and Lin (2008) developed a
directional distance function model under constant returns to scale, as follows, to
measure the efficiency of passenger production, freight production, and
consumption:
min: w1 β1 þ w2 β2 þ w3 β
Division 1
20 20
ð1Þ ð1Þ
s:t: λj X 1j ≤ ð1 - β1 ÞX 10 , λj Z 1j ≥ Z 10
j=1 j=1
20 20
ð1Þ ð1Þ
λj ðα1 X S1j Þ ≤ α1 X S10 , λj ðμ1 X S2j Þ ≤ μ1 X S20
j=1 j=1
Division 2
20 20
ð2Þ ð2Þ
λj X 2j ≤ ð1 - β2 ÞX 20 , λj Z 2j ≥ Z 20
j=1 j=1
20 20
ð2Þ ð2Þ
λj ðα2 X S1j Þ ≤ α2 X S10 , λj ðμ2 X S2j Þ ≤ μ2 X S20 ð14:3Þ
j=1 j=1
Division 3
20 20
ð3Þ ð3Þ
λj ðα3 X S1j Þ ≤ α3 X S10 , λj Z f j ≤ Z 10 , f = 1, 2
j=1 j=1
20
ð3Þ
λj Y rj ≥ ð1 þ β3 ÞY r0 r = 1, 2
j=1
i ,
αi ≤ αi ≤ α i = 1, 2, 3
i ,
μi ≤ μi ≤ μ i = 1, 2
α1 þ α2 þ α3 = 1, μ1 þ μ2 = 1
ðkÞ
λj ≥ 0, k = 1, 2, 3; j = 1, . . . , 20
where w1, w2, and w3 are the relative importance of the three divisions. A division is
efficient if the associate β value is zero, and the system is efficient if all three β values
are zero.
Although Model (14.3) is able to detect inefficiency of every division, it is not
able to show the relationship between the system and three division efficiencies. By
applying the idea of the relational model, the efficiency and three divisions can be
calculated via the following model:
E0 = max: u1 Y 10 þ u2 Y 20
s:t: v1 X 10 þ v2 X 20 þ vS1 X S10 þ vS2 X S20 = 1
w1 Z 1j - ðv1 X 1j þ vS1 α1 X S1j þ vS2 μ1 X S2j Þ ≤ 0, j = 1, . . . , 20
w2 Z 2j - ðv2 X 2j þ vS1 α2 X S1j þ vS2 μ2 X S2j Þ ≤ 0, j = 1, . . . , 20
u1 Y 1j þ u2 Y 2j - ðvS1 α3 X S1j þ w1 Z 1j þ w2 Z 2j Þ ≤ 0, j = 1, . . . , 20
α1 þ α2 þ α3 = 1, αi ≤ αi ≤ αi , i = 1, 2, 3,
μ1 þ μ2 = 1, i ,
μi ≤ μi ≤ μ i = 1, 2
u1 , u2 , v1 , v2 , vS1 , vS2 , w1 , w2 ≥ε
ð14:4Þ
At optimality, the objective value E0 is the system efficiency, and the three
division efficiencies are calculated based on the corresponding constraints, as:
u1 Y 10 þ u2 Y 20
E0 =
v1 X 10 þ v2 X 20 þ vS
1 X 10 þ v2 X 20
S S S
ð1Þ w1 Z 10
E0 =
v1 X 10 þ S S
v1 α1 X 10 þ vS S
2 μ1 X 20
ð2Þ w2 Z 20
E0 =
v2 X 20 þ S S
v1 α2 X 10 þ vS S
2 μ2 X 20
ð3Þ u1 Y 10 þ u2 Y 20

E0 =
v1 α3 X 10 þ w1 Z 10 þ w2 Z 20
S S
Comparing with the structure of the system in Fig. 14.1 where the system
efficiency is decomposable, it is noted that the structure of this problem has an
exogenous input attached to Division 3. By introducing a dummy division 4, the
original system can be transformed into a two-stage system with three divisions in
the first stage connected in parallel and one division in the second stage as depicted
in Fig. 14.2(b). The dummy division 4 has the same output as its input, which is
α3 X S1 . The function of a dummy division is to transmit information. It does not
consume, nor produce anything. The constraint corresponding to the dummy divi-
sion is redundant, and a dummy division is always efficient, with efficiency
score one.
As discussed in the previous chapters, the efficiency of a system with the
component divisions connected in series is the product of those of the component
divisions, and with the component divisions connected in parallel it is a weighted
average of those of the component divisions, where the weight associated with a
division is the proportion of the aggregate input consumed by this division in that
consumed by all divisions in the parallel structure. Based on the structure of
Fig. 14.2(b), the system efficiency can be expressed as the product of the efficiencies
of the two stages, where the efficiency of the first stage is a weighted average of the
three component divisions. In symbols, we have:
u1 Y 10 þ u2 Y 20
E0 =
v1 X 10 þ v2 X 20 þ vS

1 X 10 þ v2 X 20
S S S
vS α X S þ w1 Z 10 þ w2 Z 20 u Y 10 þ u2 Y 20

= 1 3 10 × S S1
v1 X 10 þ v2 X 20 þ v1 X 10 þ v2 X 20 v1 α3 X 10 þ w1 Z 10 þ w2 Z 20
S S S S
v X 10 þ vS S S S
1 α1 X 10 þ v2 μ1 X 20 w1 Z 10
= 1 × þ
v1 X 10 þ v2 X 20 þ vS 1 X 10 þ v2 X 20
S S S
v1 X 10 þ vS S S S
1 α1 X 10 þ v2 μ1 X 20
v2 X 20 þ vS S S S
1 α2 X 10 þ v2 μ2 X 20 w2 Z 20
× þ
v1 X 10 þ v2 X 20 þ v1 X 10 þ v2 X 20 v2 X 20 þ v1 α2 X S10 þ vS
S S S S S S
2 μ2 X 20
S
1 α3 X 10
vS vS α X S u Y 10 þ u2 Y 20

× 1S 3 10 × S S1
v1 X 10 þ v2 X 20 þ v1 X 10 þ v2 X 20 v1 α3 X 10
S S S S S
v1 α3 X 10 þ w1 Z 10 þ w2 Z 20
ð1Þ ð2Þ ð3Þ
= ωð1Þ E0 þ ωð2Þ E 0 þ ð1 - ωð1Þ - ωð2Þ Þ × 1 × E 0
where ωð1Þ = v1 X 10 þ vS S S S

1 α1 X 10 þ v2 μ1 X 20 = v1 X 10 þ v2 X 20 þ v1 X 10 þ v2 X 20
S S S S
ð2Þ S S S S
and ω = v2 X 20 þ v1 α2 X 10 þ v2 μ2 X 20 = v1 X 10 þ v2 X 20 þ v1 X 10 þ v2 X 20 :
S S S S
Note that this kind of decomposition is not applicable under variable returns to scale.
14.1.3 The Transportation Network Example
Zhao et al. (2011) proposed a distance function model to measure the efficiency of a
transportation network system composed of transportation providers (Division 1),
community (Division 2), and users (Division 3). Figure 14.3(a) shows the structure
of this system, where the transportation providers apply the operating cost (X1) to
produce revenue (Y1) as a final output, vehicle miles traveled (Z1) for community,
and average speed (Z2) as the level of service for users. The community uses vehicle
miles traveled to produce the undesirable emissions Y 2 , and the users apply the
average speed (Z2), fuel cost (X2), and travel time (X3) to produce person miles
traveled (Y3). The undesirable emissions Y 2 is superscripted with an asterisk to
denote it is an undesirable output.
The distance function model used by Zhao et al. (2011) is an output-oriented
one, with the constraints corresponding to the intermediate products expressed in
n ð1Þ n ð2Þ
equality form to maintain continuity; that is, j = 1 λj Z 1j = j = 1 λj Z 1j and
n ð1Þ n ð3Þ
j = 1 λj Z 2j = j = 1 λj Z 2j : The constraints corresponding to the three outputs
ð1Þ ð2Þ ð3Þ
are nj = 1 λj Y 1j ≥ ϕY 10 , n
j = 1 λj Y 2j ≥ Y 20 =ϕ, and
n
j = 1 λj Y 3j ≥ ϕY 30 , where
the distance parameter associated with Y 2 is 1/φ, because Y 2 is an undesirable
output, and smaller values are desired. The system efficiency is 1/φ. The division
Fig. 14.3 Structure of the a

transportation network
example. (a) Original (b)
Z1
Series-parallel conversion 2 Y2*
X1 1
3 Y3
Z2
Y1 X2 , X3
I II
Z1 Y2*
2
X1 Z1, Z 2
1
X1, X2 Y1 X2 , X3 Y3 Y1 , Y2*
3
X3 Z2 Y3
X2 , X3 X2 , X3
4
Y1 Y1
5
efficiencies in this model are not able to be measured. Moreover, the model is
nonlinear, which requires a nonlinear programming solver to find the solution. In
order to measure the division efficiencies, and decompose the system efficiency, the
original structure in Fig. 14.3(a) is converted to a series-parallel one, by introducing
a Dummy Division 4 to carry the Inputs X2 and X3 supplied from outside to Division
3, and Dummy Division 5 to carry the Output Y1 produced by Division 1, and
sending this out of the system, as shown in Fig. 14.3(b). The relational model for
measuring the system efficiency based on the converted structure is:
E 0 = max: u1 Y 10 þ u2 Y 20 þ u3 Y 30
s:t: v1 X 10 þ v2 X 20 þ v3 X 30 = 1
u1 Y 1j þ w1 Z 1j þ w2 Z 2j - v1 X 1j ≤ 0, j = 1, . . . , n
u2 Y 2j - w1 Z 1j ≤ 0, j = 1, . . . , n
ð14:5Þ
u3 Y 3j - ðv2 X 2j þ v3 X 3j þ w2 Z 2j Þ ≤ 0, j = 1, . . . , n
v2 X 2j þ v3 X 3j - ðv2 X 2j þ v3 X 3j Þ ≤ 0, j = 1, . . . , n
u1 Y 1j - u1 Y 1j ≤ 0, j = 1, . . . , n
u, vi , wg ≥ ε, r = 1, 2, 3, i = 1, 2, 3, g = 1, 2
Again, the undesirable factor Y 2 can be handled by a data transformation method.

Similar to the case discussed in Sect. 16.1, the constraints corresponding to Dummy
Divisions 4 and 5 are identities and can be deleted without affecting the solution.
After they are deleted, the resulting model is the same as that formulated based on the
original structure shown in Fig. 14.3 (a). At optimality, the system and division
efficiencies are obtained as:
u1 Y 10 þ u2 Y 20 þ u3 Y 30

E0 =
v1 X 10 þ v2 X 20 þ v3 X 30
ð1Þ u1 Y 10 þ w1 Z 10 þ w2 Z 20
E0 =
v1 X 10
ð2Þ u2 Y 20
E0 =
w1 Z 10
ð3Þ u3 Y 30
E0 =
v2 X 20 þ v3 X 30 þ w2 Z 20
Dummy Divisions 4 and 5 have a perfect efficiency score of one.

The structure of Fig. 14.3(b) is a series one, connected by two parallel sub-
systems, with components of Divisions 1 and 4 for the first, and Divisions 2, 3, and
5 for the second. The efficiency of the system can thus be expressed as the product of
those of Subsystems I and II, where the efficiency of the former is a weighted
average of those of Divisions 1 and 4, and that of the latter is a weighted average of
those of Divisions 2, 3, and 5. That is:
ð1Þ ð2Þ ð3Þ

E 0 = ωð1Þ E 0 þ ωð4Þ × ωð2Þ E0 þ ωð3Þ E 0 þ ωð5Þ
where ωð1Þ = v1 X 10 = v1 X 10 þ v2 X 20 þ v3 X 30 , ωð2Þ = w1 Z 10 = v2 X 20 þ v3 X 30

þw1 Z 10 þ w2 Z 20 þ u1 Y 10 Þ,ω = v2 X 20 þ v3 X 30 þ w2 Z 20 =
ð3Þ
v2 X 20 þ v3 X 30 þ w1 Z 10 þ w2 Z 20 þ u1 Y 10 , ω(4) = 1 - ω(1), and ω(5) = 1 -

ω(2) - ω(3). For divisions in the same parallel structure, the one with a larger weight
has a greater impact on the system’s performance.
14.1.4 The Bank Profit Centers Example
Avkiran (2009) applied a slacks-based model to measure the performance of fifteen

UAE (United Arab Emirates) banks, in which three profit centers were involved:
loans, advances, and overdrafts (Division 1), mortgaged real estate loans (Division
2), and discounted commercial bills (Division 3), with the structure shown in
Fig. 14.4(a). In this system Division 1 applies interest expenses (X1) and
non-interest expenses (X2) to generate interest income (Y1) and non-interest income

bank profit centers example.
(a) Original (b) Series- Y1 , Y2
parallel conversion
X1, X 2
1
Z Y5 ,Y6 Yr
Xi X5, X6 3
i = 1,...,6 r = 1,...,6
X 3, X 4 Y3 ,Y4
2
I II
X1, X 2 Y1 , Y2
1
Z X5, X6 Y5 , Y6
3
Xi X 3, X 4 Y3 , Y4 Z Yr
2
i = 1,...,6 r =1,...,6
Yr Yr
5 r =1,...,4
X5, X6 X5, X6 r = 1,...,4
4
(Y2) as outputs, and non-interest income as intermediate product (Z ) for Division 3 to

use. Note that the same non-interest income is separated into Y2 and Z for different
uses. Division 2 also consumes interest and non-interest expenses to generate
incomes. Since the expenses are not shared with Division 1, and neither are the
incomes, we use X3 and X4 to express the inputs, and Y3 and Y4 to express outputs of
this division. Similarly, Division 3 also consumes interest expenses (X5) and
non-interest expenses (X6) to generate interest income (Y5) and non-interest income
(Y6). This division also consumes Z, generated from Division 1, to carry out its
operations.
The conventional way of using the ratio of the average input distance parameter to
that of the output distance parameter to represent the efficiency of a division is
ð3Þ
adopted by Avkiran (2009). For example, the efficiency of Division 3 is ρ0 =
1 - 12 s5- =X 50 þ s6- =X 60 = 1 þ 12 sþ þ - þ
5 =Y 50 þ s6 =Y 60 , where si and sr are the
slack variables associated with Input Xi and Output Yr, respectively. The system
efficiency is defined as the ratio of the weighted average of the input distance
parameters of the three divisions to that of the output distance parameters, following
the idea in Tone and Tsutsui (2009). The system efficiency calculated this way
cannot be expressed by the division efficiencies. However, the relational model can
achieve this, as explained below.
This system is actually a parallel one, connected by Division 2 and a series
subsystem composed of Divisions 1 and 3. Since Division 1 produces outputs that
are sent to two destinations, and Division 3 receives inputs from two sources, two
Dummy Divisions 4 and 5 are needed to carry the exogenous inputs required by
Division 3 and the final outputs produced by Divisions 1 and 2, respectively, with the
structure shown in Fig. 14.4(b). Based on this structure, the relational model for this
system under constant returns to scale is:
6
E 0 = max: ur Y r0
r=1
6
s:t: vi X i0 = 1
i=1
u1 Y 1j þ u2 Y 2j þ wZ j - ðv1 X 1j þ v2 X 2j Þ ≤ 0, j = 1, . . . , 15
u3 Y 3j þ u4 Y 4j - ðv3 X 3j þ v4 X 4j Þ ≤ 0, j = 1, . . . , 15
u5 Y 5j þ u6 Y 6j - ðv5 X 5j þ v6 X 6j þ wZ j Þ ≤ 0, j = 1, . . . , 15
v5 X 5j þ v6 X 6j - ðv5 X 5j þ v6 X 6j Þ ≤ 0, j = 1, . . . , 15
4 4
r = 1 ur Y r0 - r = 1 ur Y r0 ≤ 0, j = 1, . . . , 15
ur , vi , w ≥ ε, r = 1, . . . , 6, i = 1, 2, 3
ð14:6Þ
The constraints corresponding to Dummy Divisions 4 and 5 are identities and can
be deleted without affecting the solution. After they are deleted, the resulting model
is the same as that formulated from the original structure.
At optimality, the system and division efficiencies, based on the constraints, are
calculated as:
6
r = 1 ur Y r0
E0 = 6
i = 1 vi X i0
ð1Þ u1 Y 10 þ u2 Y 20 þ w Z 0

E0 =
v1 X 10 þ v2 X 20
ð2Þ u3 Y 30 þ u4 Y 40
E0 =
v3 X 30 þ v4 X 40
ð3Þ u5 Y 50 þ u6 Y 60
E0 =
v5 X 50 þ v6 X 60 þ w Z 0

The structure in Fig. 14.4(b) shows that the system is a series one, connected by
two parallel subsystems, one with components of Divisions 1, 2, and 4, and the other
with components of Divisions 3 and 5. The system efficiency can thus be
expressed as:
14.2 Four-Division Structures 375
ðIÞ ðIIÞ
E0 = E0 × E0
ð1Þ ð2Þ ð3Þ
= ωð1Þ E 0 þ ωð2Þ E0 þ ωð4Þ × ωð3Þ E 0 þ ωð5Þ
where ωð1Þ = 2i = 1 vi X i0 = 6i = 1 vi X i0 , ωð2Þ = 4

i = 3 vi X i0 =
6
i = 1 vi X i0 , ω
(4)
=1 -
ω(1) -ω(2), ωð3Þ = 6
i = 5 vi X i0 þ w Z 0 =
6
i = 5 vi X i0 þ w Z 0 þ
4
r = 1 ur Y r0 ,
and ω(5) = 1 - ω(2).
14.2 Four-Division Structures
Network structures with four divisions are also common in the literature, although
they are not as frequent as three-division structures. We introduce four examples
from the literature.
14.2.1 The International Tourist Hotel Example
Many organizations have several functions, where different inputs are applied to
produce different outputs. For example, an international tourist hotel usually has two
functions, accommodations and catering. These two functions operating indepen-
dently, and can be expressed as two parallel subsystems. Similar to the railway
service example discussed in Sect. 14.1.2, the services provided for these two
functions are lost if they are not consumed. In this regard, the operations of each
function are separated into two stages, where services are provided in the first stage
and consumed in the second stage. Hsieh and Lin (2010) developed a relational
model to measure the efficiency of 57 international tourist hotels in Taiwan. For
accommodations, the accommodations cost (X1) and accommodations employees
(X2) are used to provide room services (Z1) in the accommodations production
process (Division 1). The rooms are then used by customers to generate accommo-
dations revenue (Y1) in the accommodations service process (Division 3). Similarly,
the catering cost (X3) and catering employees (X4) are used to provide catering floors
(Z2) in the catering production process (Division 2). The catering floors are then used
to generate catering revenue (Y2) in the catering service process (Division 4).
Figure 14.5(a) depicts the structure of this problem. Hsieh and Lin (2010) developed
a relational model, under constant returns to scale, to calculate the system and
division efficiencies:

international tourist hotels
example. (a) Original (b) X1 Z1 Y1
Series conversion (c) 1 3
Parallel conversion X2
X3 Z2 Y2
2 4
X4
X1 Z1 Z1 Y1
1 3
X 1, X 2 X2
Y1
X 3, X 4 Y2
X3 Z2 Z2 Y2
2 4
X4
X1 Z1 Y1
1 3
X 1, X 2 X2
Y1
X 3, X 4 Y2
X3 Z2 Y2
2 4
X4
E 0 = max: u1 Y 10 þ u2 Y 20
4
s:t: vi X i0 = 1
i=1
w1 Z 1j - ðv1 X 1j þ v2 X 2j Þ ≤ 0, j = 1, . . . , 57
ð14:7Þ
w2 Z 2j - ðv3 X 3j þ v4 X 4j Þ ≤ 0, j = 1, . . . , 57
u1 Y 1j - w1 Z 1j ≤ 0, j = 1, . . . , 57
u2 Y 2j - w2 Z 2j ≤ 0, j = 1, . . . , 57
u1 , u2 , vi , w1 , w2 ≥ ε, i = 1, 2, 3, 4
After a set of optimal solutions is obtained, the system and division efficiencies,
based on the constraints, are calculated as:
u1 Y 10 þ u2 Y 20
E0 = 4
i = 1 vi X i0
ð1Þ w1 Z 10
E0 =
v1 X 10 þ v2 X 20
ð2Þ w2 Z 20
E0 =
v3 X 30 þ v4 X 40
ð3Þ u1 Y 10
E0 =
w1 Z 10
ð4Þ u2 Y 20
E0 =
w2 Z 20
To express the system efficiency by division efficiencies, Hsieh and Lin (2010)
transformed the structure of the problem into a two-stage structure, where the first
stage is composed of Divisions 1 and 2 connected in parallel and the second stage is
composed of Divisions 3 and 4 also connected in parallel, as shown in Fig. 14.5(b).
Based on the efficiency decomposition for series and parallel structures, the
system efficiency can be decomposed into the product of the two stage efficiencies,
where the first stage efficiency is a weighted average of Division 1 and Division
2 efficiencies and the second stage efficiency is a weighted average of Division 3 and
Division 4 efficiencies. In symbols, it is:

E 0 = ωð1Þ E 0 þ ωð2Þ E0 × ωð3Þ E 0 þ ωð4Þ E 0
where ωð1Þ = 2i = 1 vi X i0 = 4i = 1 vi X i0 , ω(2) = 1 - ω(1), ωð3Þ = w1 Z 10 = 2g = 1 wg Z g0 ,

and ω(4) = 1 - ω(3). It should be noted that the way that the efficiency of a system is
decomposed is not unique. For example, the structure of this problem can also be
converted to a parallel structure composed of two subsystems, where each subsystem
is composed of two divisions connected in series, as shown in Fig. 14.5(c). Based on
this structure, the system efficiency is decomposed as:

E0 = ωð1Þ E 0 × E0 þ ωð2Þ E 0 × E0
where ωð1Þ = 2
i = 1 vi X i0 =
4
i = 1 vi X i0 and ω(2) = 1 - ω(1).
14.2.2 The Environmental Protection Example
Huang et al. (2014) proposed a slacks-based model to measure the efficiency of the
environmental protection system for 22 municipalities in Taiwan that include three
stages: administration, execution, and environmental protection, where the second
stage consists of two processes, waste treatment, and environmental auditing, with
the structure shown in Fig. 14.6. There are four divisions in this system, where
Division 1 (administration) converts government expenditure (X) into garbage

environmental protection Z1, Z 2 Z 4 , Z5
example 2
X 1 4 Y1* , Y2*
Z3 Z6, Z7
3
trucks (Z1), treatment workers (Z2), and investigation workers (Z3). Division 2 (waste
treatment) then uses Z1 and Z2 to generate incinerated and buried garbage (Z4) and
recycled garbage (Z5), and Division 3 (environmental auditing) uses Z3 to generate
water examinations (Z6) and examined factories (Z7). Finally, Division 4 (environ-
mental protection) applies Z4, Z5, Z6, and Z7 to produce the undesirable outputs of
ozone and sulfur dioxide (Y 1 ) and water pollution (Y 2 ). The undesirable factors are
superscripted with asterisks.
For the Intermediate Products Z4, Z5, Z6, and Z7, Huang et al. (2014) on the one
hand treated these as desirable outputs of Divisions 2 and 3, and on the other hand
treated them as undesirable inputs of Division 4. The model for measuring the
efficiency of this system was formulated as:

min: ρ0 þ ρ0 þ ρ0 þ ρ0
22
ð1Þ
s:t: λj X j þ s - = X 0
j=1
22
ð1Þ
λj Z gj - sþ
g = Z g0 , g = 1, 2, 3
j=1
22
ð2Þ
λj Z gj þ sg- = Z g0 , g = 1, 2
j=1
22
ð2Þ
λj Z gj - sþ
g = Z g0 , g = 4, 5
j=1
22
ð14:8Þ
ð3Þ
λj Z 3j þ s3- = Z 30
j=1
22
ð3Þ
λj Z gj - sþ
g = Z g0 , g = 6, 7
j=1
22
ð4Þ
λj Z gj - ~sg- = Z g0 , g = 4, 5, 6, 7
j=1
22
ð4Þ
λj Y rj þ ~sþ
r = Y r0 , r = 1, 2
j=1
ðkÞ
λj ≥ 0, k = 1, 2, 3, 4; j = 1, . . . , 22
where
ð1Þ 1 - s - =X 0
ρ0 =
1 þ
1 þ s1 =Z 10 þ sþ þ
2 =Z 20 þ s3 =Z 30
3
1
ð2Þ
1 - s1- =Z 10 þ s2- =Z 20
ρ0 = 2
1
1 þ sþ =Z þ sþ 5 =Z 50
2 4 40 ð14:9Þ
-
ð3Þ 1 - s3 =Z 30
ρ0 =
1 þ
1 þ s6 =Z 60 þ sþ 7 =Z 70
2
1 þ
ð4Þ
1 - s1 =Y 10 þ sþ
1 =Y 20
ρ0 = 2
1
1 þ s4- =Z 40 þ s5- =Z 50 þ s6- =Z 60 þ s7- =Z 70
4
ð2Þ ð3Þ
Note that in calculating ρ0 and ρ0 the Intermediate Products Z4, Z5, Z6, and Z7
ð4Þ
were treated as desirable factors, while in calculating ρ0 they were treated as
undesirable factors. The original model assumes variable returns to scale, with the
22 ðk Þ ðk Þ
constraints j = 1 λj = 1, k = 1, . . ., 4, added. At optimality, ρ0 , defined in
Eq. (14.9), are the efficiencies of the four divisions.
Model (14.8) is nonlinear. Moreover, it is not able to measure the efficiency of the
system, although the objective function implicitly implies that the system efficiency
is the average of the four division efficiencies. It would thus be desirable to have a
linear model, and obtain a relationship between the system and division efficiencies
which is reflected in the structure of the system. Applying the relational model
proposed by Kao (2009), the system efficiency, based on the structure of Fig. 14.6,
can be formulated as:
u1 Y 10 þ u2 Y 20
E0 = max:
vX 0
s:t: w1 Z 1j þ w2 Z 2j þ w3 Z 3j - vX j ≤ 0, j = 1, . . . , 22
w4 Z 4j þ w5 Z 5j - ðw1 Z 1j þ w2 Z 2j Þ ≤ 0, j = 1, . . . , 22
w6 Z 6j þ w7 Z 7j - w3 Z 3j ≤ 0, j = 1, . . . , 22
u1 Y 1j þ u2 Y 2j - ðw4 Z 4j þ w5 Z 5j þ
w6 Z 6j þ w7 Z 7j Þ ≤ 0, j = 1, . . . , 22
ur , v, wg ≥ ε, r = 1, 2; g = 1, . . . , 7 ð14:10Þ
The undesirable factors can be handled by the data transformation methods

discussed in Chap. 6. At optimality, the system and division efficiencies are
calculated as:
u1 Y 10 þ u2 Y 20

E0 =
v X 0
ð1Þ w1 Z 10 þ w2 Z 20 þ w3 Z 30
E0 =
v X 0
ð2Þ w4 Z 40 þ w5 Z 50
E0 =
w1 Z 10 þ w2 Z 20
ð3Þ w6 Z 60 þ w7 Z 70
E0 =
w3 Z 30
ð4Þ u1 Y 10 þ u2 Y 20
E0 =
w4 Z 40 þ w5 Z 50 þ w6 Z 60 þ w7 Z 70
A closer examination of the structure in Fig. 14.6 reveals that it is a series one,
connected by Division 1, Division 4, and a parallel subsystem composed of
Divisions 2 and 3. Based on the decomposition principle for series and parallel
structures, the system efficiency can be expressed as:

E 0 = E0 × ωð2Þ E 0 þ ωð3Þ E0 × E0
where ωð2Þ = w1 Z 10 þ w2 Z 20 = w1 Z 10 þ w2 Z 20 þ w3 Z 30 and ωð3Þ =

ð2Þ
w3 Z 30 = w1 Z 10 þ w2 Z 20 þ w3 Z 30 Þ = 1 - ω : This relationship shows that Divisions
1 and 4 have greater impacts on the performance of the system than Divisions 2 and
3, and the values of ω(2) and ω(3) indicate the division that has a larger impact on the
performance of the second stage.
14.2.3 The Corporate and Consumer Banking Example
Lin and Chiu (2013) measured the performance of 30 Taiwanese banks by separating
their operations into three stages: production, service, and profitability, where the
service stage was further separated into corporate banking and consumer banking,
with the structure shown in Fig. 14.7(a). In this system, Division 1 (production) uses
fixed assets (X1), operating expenses (X2), and equity (X3) to generate corporate
banking loans (Z1), consumer banking loans (Z2), and deposits (Z3). The perfor-
mance of this division is termed production efficiency. Division 2 (corporate bank-
ing) then converts Z1 into collected corporate banking loans (Z4), and Division
3 (consumer banking) converts Z2 into collected consumer banking loans (Z5). The
performance of this stage is termed service efficiency. Finally, Division

corporate and consumer
banking example. (a) Z1 Z4
Original (b) Series-parallel 2
conversion
X1, X 2 Z2 Z5 Y1, Y2
1 3 4
X3 Y3
Z3
I
II
Z1 Z4
2
Y1 , Y2
X1, X 2 4
1
Z2 Z5 Y3
X3 3
Z3 Z3
5
4 (profitability) uses Z3, Z4, and Z5 to generate interest revenue (Y1), fee revenue (Y2),
and profit (Y3). The performance of this stage is termed profitability efficiency.
The division efficiency was defined in the conventional way of taking the ratio of
the average input distance parameter to the average output distance parameter, and
the system efficiency was defined as the ratio of the weighted average of the input
distance parameters of the four divisions to that of the output distance parameters
(Tone & Tsutsui, 2009). The system efficiency in this case cannot be expressed by
the division efficiencies with a specific mathematical equation. In order to find the
relationship between the system and division efficiencies, the relational model of
Kao (2009) is thus applied.
The structure of the system is first converted to a series-parallel one by introduc-
ing the Dummy Division 5, as depicted in Fig. 14.7(b). The relational model for this
structure under constant returns to scale is:
3
r = 1 ur Y r0
E 0 = max: 3
i = 1 vi X i0
3 3
s:t: wg Z gj - vi X ij ≤ 0, j = 1, . . . , 30
g=1 i=1
w4 Z 4j - w1 Z 1j ≤ 0, j = 1, . . . , 30
ð14:11Þ
w5 Z 5j - w2 Z 2j ≤ 0, j = 1, . . . , 30
3 5
ur Y rj - wg Z gj ≤ 0, j = 1, . . . , 30
r=1 g=3
w3 Z 3j - w3 Z 3j ≤ 0, j = 1, . . . , 30
ur , vi , wg ≥ ε, r = 1, 2, 3, i = 1, 2, 3, g = 1, . . . , 5
If we delete the redundant identity constraint corresponding to Dummy Division

5, we obtain a model which is the same as that formulated based on the original
structure. At optimality, we have the following system and division efficiencies:
3
r = 1 ur Y r0
E0 = 3
i = 1 vi X i0
3
ð1Þ g = 1 wg Z g0
E0 = 3
i = 1 vi X i0
ð2Þ w4 Z 40
E0 =
w1 Z 10
ð3Þ w5 Z 50
E0 =
w2 Z 20
3
ð4Þ r = 1 ur Y r0
E0 = 5
g = 3 wg Z g0
The structure of Fig. 14.7(b) shows that this system has a series structure,
connected by Division 1, Subsystem I, and Division 4. Of which, Subsystem I is a
parallel one, connected by Division 5 and a parallel Subsystem II, composed of
Divisions 2 and 3. Based on the structure shown in Fig. 14.7 (b), the system
efficiency can be expressed as:
ð1Þ ðIÞ ð4Þ ð1Þ ðIIÞ ð4Þ

E 0 = E0 × E 0 × E 0 = E0 × ωðIIÞ E 0 þ ωð5Þ × E0
= E0 × ωðIIÞ ωð2Þ E 0 þ ωð3Þ E0 þ ωð5Þ × E0
where ωðIIÞ = w1 Z 10 þ w2 Z 20 = w1 Z 10 þ w2 Z 20 þ w3 Z 30 , ωð5Þ =

w3 Z 30 = w1 Z 10 þ w2 Z 20 + w3 Z 30 Þ = 1 - ω , ω = w1 Z 10 / w1 Z 10 + w2 Z 20 Þ,
(II) (2)
and ω(3) = w2 Z 20 = w1 Z 10 þ w2 Z 20 = 1 - ωð2Þ : The production, service, and
ð1Þ ðIIÞ
profitability efficiencies are those of Division 1 (E 0 ), Subsystem II (E 0 ), and
ð4Þ
Division 4 (E0 ), respectively. Based on the definitions of ω(II), ω(2), and ω(3), the
expression of E0 can be simplified slightly to:

E0 = E 0 × ωð2Þ E0 þ ωð3Þ E 0 þ ωð 5 Þ × E 0
where ωð2Þ = w1 Z 10 = w1 Z 10 þ w2 Z 20 þ w3 Z 30 and ωð3Þ =

w2 Z 20 = w1 Z 10 þ w2 Z 20 þ w3 Z 30 Þ:
14.2.4 The Matrix Structure Example
Cheng and Gao (2015) used the conventional ratio-form model to measure the
efficiency of systems with matrix structures. The efficiency of the division was the
ratio of the aggregation of the outputs, including both exogenous and intermediate,
to that of the inputs, including also both exogenous and intermediate. The system
efficiency was defined as a weighted aggregation of the division efficiencies, where
the weight associated with a division was the proportion of the aggregate input
consumed by this division in that consumed by all divisions. The model was used to
measure the efficiency of the secondary industry of 27 provinces in China, where the
system is composed of four sectors of mining (Division 1), electricity, gas, and water
production (Division 2), manufacturing (Division 3), and construction (Division 4).
The system has the structure shown in Fig. 14.8.
Fig. 14.8 A system with Y1 Y3

the matrix-type structure
Z3
X1 1 3 X3
Z4
Z5
Z8
Z1 Z 2 Z11 Z12
Z7
Z6
Z9
X2 2 4 X4
Z 10
Y2 Y4
Based on this structure, the following ratio-form model was used to measure the
system efficiency under constant returns to scale:
4
r = 1 ur Y r0
E0 = max: 4
i = 1 vi X i0
s:t: ðu1 Y 1j þ w1 Z 1j þ w3 Z 3j þ w5 Z 5j Þ -
ðv1 X 1j þ w2 Z 2j þ w4 Z 4j þ w6 Z 6j Þ ≤ 0, j = 1, . . . , 27
ðu2 Y 2j þ w2 Z 2j þ w7 Z 7j þ w9 Z 9j Þ -
ur , vi , wg ≥ ε, r = 1, . . . , 4, i = 1, . . . , 4, g = 1, . . . , 12
ð14:12Þ
At optimality, the system and division efficiencies are calculated as:
4
r = 1 ur Y r0
E0 = 4
i = 1 vi X i0
ð1Þ u1 Y 10 þ w1 Z 10 þ w3 Z 30 þ w5 Z 50

E0 =
v1 X 10 þ w2 Z 20 þ w4 Z 40 þ w6 Z 60
ð2Þ u2 Y 20 þ w2 Z 20 þ w7 Z 70 þ w9 Z 90
E0 =
v2 X 20 þ w1 Z 10 þ w8 Z 80 þ w10 Z 10,0
ð3Þ u3 Y 30 þ w4 Z 40 þ w8 Z 80 þ w11 Z 11,0
E0 =
v3 X 30 þ w3 Z 30 þ w7 Z 70 þ w12 Z 12,0
ð4Þ u4 Y 40 þ w6 Z 60 þ w10 Z 10,0 þ w12 Z 12,0
E0 =
v4 X 40 þ w5 Z 50 þ w9 Z 90 þ w11 Z 11,0
For systems with feedbacks, it is impossible to convert the system into a series-
parallel structure by introducing dummy divisions. In this example, every two
divisions communicate with each other, which is a case of feedback, the system
efficiency cannot be decomposed into the division efficiencies. To develop a rela-
tionship between the system and division efficiencies in this case, the authors applied
the idea of additive efficiency aggregation in Cook et al. (2010) to define the system
14.3 Five-Division Structures 385
efficiency as a weighted average of the division efficiencies. Let ωð1Þ =

v1 X 10 þ w2 Z 20 þ w4 Z 40 þ w6 Z 60 =T, ωð2Þ = v2 X 20 þ w1 Z 10 þ w8 Z 80 þ
w10 Z 10,0 Þ=T, ω = v3 X 30 þ w3 Z 30 þ w7 Z 70 þ w12 Z 12,0 =T, and ωð4Þ = v4 X 40 þ
ð3Þ
w5 Z 50 þ w9 Z 90 þ w11 Z 11,0 Þ=T, where T = 4i = 1 vi X i0 þ 12

g = 1 wg Z g0 : The
weighted average of the division efficiencies becomes:
4 12
r = 1 ur Y r0 þ
4
ðk Þ g = 1 wg Z g0
ωðkÞ E 0 = 4 12
i = 1 vi X i0 þ

k=1 g = 1 wg Z g0
And the aggregation model used by the authors for this system is:
4 12
r = 1 ur Y r0 þ g = 1 wg Z g0
E 0 = max: 4
i = 1 vi X i0 þ
12
g = 1 wg Z g0
ð14:13Þ
s:t: The constraints of Model ð14:12Þ
This model can be linearized by applying the variable substituting technique of

Charnes and Cooper (1962). After the optimal solution is obtained, the values of the
weights show which divisions have a greater impact on the performance of the
system. The difference between this model and Model (14.12) is that the former
includes the intermediate products in both the numerator and denominator of the
fractional objective function in defining the system efficiency. Since the system
efficiency 4r = 1 ur Y r0 = 4i = 1 vi X i0 defined in Model (14.12) is less than or equal to
4 12 4
one, it will be less than or equal to r = 1 ur Y r0 þ g = 1 wg Z g0 = i = 1 vi X i0 þ
12
g = 1 wg Z g0 Þ, the system efficiency defined in Model (14.13).
14.3 Five-Division Structures
In the real world, a network system can have a large number of divisions. However,
it is usually simplified by aggregating relatively less important divisions to make the
analysis possible and easier. Excluding theoretical researches, five divisions are
probably the largest number of divisions that have been reported in the literature.
In this section, we discuss two examples of five divisions.
14.3.1 The Major League Baseball Example
Lewis and Sexton (2004) studied an interesting problem of measuring the efficiency
of the 30 baseball teams who play Major League Baseball (MLB) in the US over one
season. Each team is considered as a DMU, with five divisions, as depicted in

Fig. 14.9(a). Divisions 1 and 2 constitute the front office operations and Divisions
3, 4, and 5 constitute the on-field operations. Among the five divisions, Division
1 spends money on position player salaries (X1) to provide the team with offensive
production in terms of total bases gained (Z1) and walks gained (Z2), while surren-
dering defensive errors (Z 3 ). The errors surrendered are undesirable outputs, and
are marked with an asterisk. Division 2 spends money on pitcher salaries (X2),
and pitchers provide the team with defensive production in terms of total bases
surrendered (Z 4 ) and walks surrendered (Z 5 ). Here both total bases surrendered and
walks surrendered are undesirable outputs. Division 3 converts total bases gained
(Z1), walks gained (Z2), and errors gained (X3), as committed by the team’s oppo-
nents, into runs gained (Z6). Division 4 converts errors surrendered (Z 3 ), total
bases surrendered (Z 4 ), and walks surrendered (Z 5 ) into runs surrendered (Z 7 ).
Obviously, runs surrendered are undesirable. Finally, Division 5 converts
runs gained (Z6) and runs surrendered (Z 7 ) into games won (Y ). The efficiency of
each division was measured independently using a distance function model.
For undesirable factors, the signs of the observations are reversed, and the
corresponding distance parameters use reciprocals. Details of this idea are described
in Chap. 6.
To measure the efficiency of this system, a distance function model was first
applied to Divisions 1 and 2 to find the target values of the Intermediate Products Z1,
Z2, Z 3 , Z 4 , and Z 5 that these two divisions need to meet to become efficient. These
target values, together with other inputs and outputs, were then used to calculate the
efficiencies of Divisions 3 and 4, which would yield the target values for the
Intermediate Products Z6 and Z 7 at the same time. Finally, the distance function
model was applied to Division 5, using these target values, to find the target value for
the output Y, and the ratio of the observed Y to the target Y is the efficiency of the
system. Obviously, the system will be inefficient if any of the divisions is inefficient.
However, how the system efficiency is related to the division efficiencies is not clear,
and it would be helpful for management purposes if this relationship were known.
Moreover, if there is more than one output, then the method used by Lewis and
Sexton (2004) encounters difficulties with regard to aggregating the efficiencies of
different outputs together. For these reasons, we need a more suitable model for
efficiency measurement.
By adding one Dummy Division 6 to carry the input supplied from outside to
Division 3, the original structure can be converted to a series structure connected by
Subsystems I and II, and Division 5, where Subsystem I is composed of Divisions
1, 2, and 6, connected in parallel, and Subsystem II is composed of Divisions 3 and
4, also connected in parallel, as shown in Fig. 14.9(b). Based on this structure, the
system efficiency, under constant returns to scale, can be measured via the following
ratio-form model:
Fig. 14.9 Structure of the a X3

MLB baseball example. (a)
Original (b) Series-parallel
conversion Z1 , Z 2 Z6
X1 1 3
Z 3*
5 Y
Z 4* , Z 5* Z 7*
X2 2 4
b
I II
X1 Z1 , Z 2
1
Z 3* Z1 , Z 2 Z6
3
X1, X 2 , X 3 X2 Z ,Z
* * X3
2 4 5 Y
5
Z *
3
Z *
7
X3 X3
4
6 Z 4* , Z5*
uY 0
E0 = max: 3
vi X i0
i=1
s:t: w1 Z 1j þ w2 Z 2j þ w3 Z 3j - v1 X 1j ≤ 0, j = 1, . . . , 30
w4 Z 4jþ w5 Z 5j
- v2 X 2j ≤ 0, j = 1, . . . , 30
w6 Z 6j - ðv3 X 3j þ w1 Z 1j þ w2 Z 2j Þ ≤ 0, j = 1, . . . , 30
w7 Z 7j - ðw3 Z 3j þ w4 Z 4j þ w5 Z 5j Þ ≤ 0, j = 1, . . . , 30
uY j - ðw6 Z 6j þ w7 Z 7j Þ ≤ 0, j = 1, . . . , 30
v3 X 3j - v3 X 3j ≤ 0, j = 1, . . . , 30
u, vi , wg ≥ ε, i = 1, 2, 3, g = 1, . . . , 7
ð14:14Þ
The undesirable factors can be handled by the data transformation method

presented in Chap. 6. In this model the objective function is the system efficiency,
and each constraint corresponds to the operation of one division. One point to be
noted is that the constraint corresponding to the dummy division, v3X3j- v3X3j ≤ 0,
is an identity, which always holds, and can be deleted without affecting the optimal
solution. If it is deleted, then the model is the same as that formulated from the
original structure. Another point to be noted is that the dummy division is always
efficient because its input and output are the same.
After a set of optimal solutions (u*, v*, w*) is obtained, the system and division
efficiencies, based on the constraints, are obtained as:
u Y 0
E0 =
v1 X 10 þ v2 X 20 þ v3 X 30
ð1Þ w1 Z 10 þ w2 Z 20 þ w3 Z 30
E0 =
v1 X 10
ð2Þ w4 Z 40 þ w5 Z 50
E0 =
v2 X 20
ð3Þ w6 Z 60
E0 =
v3 X 30 þ w1 Z 10 þ w2 Z 20
ð4Þ w7 Z 70
E0 =
w3 Z 30 þ w4 Z 40 þ w5 Z 50
ð5Þ u Y 0
E0 =
w6 Z 60 þ w7 Z 70
Based on the structure of Fig. 14.9(b), the efficiency of the system can be
decomposed as the product of those of Subsystem I, Subsystem II, and Division
5, and the efficiencies of the two subsystems are weighted averages of those of
Divisions 1, 2, and 6, and Divisions 3 and 4, respectively. In symbols, we have:

E 0 = ωð1Þ E0 þ ωð2Þ E0 þ ωð6Þ × ωð3Þ E 0 þ ωð4Þ E0 × E0
where ωð1Þ = v1 X 10 = v1 X 10 þ v2 X 20 þ v3 X 30 , ωð2Þ =

v2 X 20 = v1 X 10 þ v2 X 20 þ v3 X 30 , ω(6)= 1 - ω(1) - ω(2), ωð3Þ =
v3 X 30 þ w1 Z 10 þ w2 Z 20 = v3 X 30 þ 5
g = 1 wg Z g0 , and ω(4) = 1 - ω(3).
Intuitively, the components of a series structure have a greater impact on system
performance than that of a parallel structure. To investigate this property, suppose
every division has the same efficiency score E, and the weights for the divisions in
the same parallel structure are the same. Increasing the efficiency score of Division
5, a division in the series structure, by δ increases the efficiency of the system by
3 E þ 3 × ðE Þδ = δ × 3 E þ 3 E . However, increasing the efficiency score of
2 1 2 2 1
Division 1, a division in a parallel structure, by δ increases the efficiency of the

system by 13 δ × ðE Þ × E = δ × 13 E 2 , which is obviously smaller than
δ × 23 E 2 þ 13 E . Another intuition is that a division in a parallel structure with
fewer divisions has a greater impact on the system performance than that of a
parallel structure with more divisions. For example, increasing the efficiency score
of Division 3, a division in the two-division parallel structure, by δ increases the
efficiency of the system by 23 E þ 13 × 12 δ × E = δ × 13 E 2 þ 16 E , which is greater

than δ × 13 E 2 , the increase in the system efficiency caused by increasing the
efficiency of Division 1, a division in the three-division parallel structure, by δ.
For divisions in the same parallel structure, the actual values of the associated
weights determine the ones that have greater impacts on the system efficiency.
14.3.2 The NBA Basketball Example
Following the idea of Lewis and Sexton (2004) for evaluating the performance of
30 baseball teams in the MLB, Moreno and Lozano (2014) evaluated the perfor-
mance of 30 NBA basketball teams for the 2009–2010 season using a slacks-based
model. The system consists of five divisions: team-work performance (Division 1),
first team offensive (Division 2), defensive (Division 3), bench team offensive
(Division 4), and win generation (Division 5), with the structure shown in
Fig. 14.10. The interdependence of the divisions in this system is that Division
1 uses the first team budget (X1) and bench team budget (X2) to generate two-point
shots by the first team (Z1) and bench team (Z13), three-point shots by the first team
(Z2) and bench team (Z14), free throws by the first team (Z3) and bench team (Z15),
offensive rebounds by the first team (Z4) and bench team (Z16), assists by the first
team (Z5) and bench team (Z17), and the inverse of turnovers by the first team (Z6)
and bench team (Z18), defensive rebounds by the first team (Z7) and bench team
(Z10), steals by the first team (Z8) and bench team (Z11), and blocked shots by the first
team (Z9) and bench team (Z12). Division 2 then uses Z1 ~ Z6 to generate points by
the first team (Z19), Division 3 uses Z7 ~ Z12 to generate the inverse of points by the
opponent (Z20), and Division 4 uses Z13 ~ Z18 to generate points by the bench team
(Z21). Finally, Division 5 aggregates Z19 ~ Z21 to yield the number of team victories
(Y ).
Since the budget was assumed to be relocatable, the constraint corresponding to
ð1Þ
the two inputs is 30
j = 1 λj X 1j þ X 2j þ s - = X 10 þ X 20 : The model used by
Moreno and Lozano (2014) is:

NBA basketball example Zg Z19
2
g = 1,...,6
X1, X 2 Zg Z 20
Y
1 3 5
g = 7,...,12
Zg Z 21
4
g = 13,...,18
1 - s - =ðX 10 þ X 20 Þ
ρ0 = max:
1 þ sþ =Y 0
s:t: The conventional constraints of the slacksbased model:
ð14:15Þ
where s+ is the slack variable associated with the output Y, defined as

30 ð5Þ þ
j = 1 λj Y j - s = Y 0 : Although the efficiency of each division can be measured
by applying the conventional definition of the ratio of the average input distance
parameter to the average output distance parameter, the relationship between the
system and division efficiencies is still not known. A relational model is thus used to
obtain the relationship.
Based on the inputs and outputs of every division, the relational model of this
system can be formulated as:
uY 0
E 0 = max:
vðX 10 þ X 20 Þ
18
s:t: wg Z gj - vðX 1j þ X 2j Þ ≤ 0, j = 1, . . . , 30
g=1
6
w19 Z 19j - wg Z gj ≤ 0, j = 1, . . . , 30
g=1
12
ð14:16Þ
w20 Z 20j - wg Z gj ≤ 0, j = 1, . . . , 30
g=7
18
w21 Z 21j - wg Z gj ≤ 0, j = 1, . . . , 30
g = 13
21
uY j - wg Z gj ≤ 0, j = 1, . . . , 30
g = 19
u, v, wg ≥ ε, g = 1, . . . , 21
After an optimal solution is obtained, the system and division efficiencies are
calculated as:
u Y 0
E0 =
v ð X 10 þ X 20 Þ
18
ð1Þ g = 1 wg Z g0
E0 =
v ðX 10 þ X 20 Þ
ð2Þ w19 Z 19,0
E0 = 6
g = 1 wg Z g0
ð3Þ w20 Z 20,0

E0 = 12
g = 7 wg Z g0
ð4Þ w21 Z 21,0

E0 = 18
g = 13 wg Z g0
ð5Þ u Y 0
E0 = 21
g = 19 wg Z g0
To decompose the system efficiency, it is noted that the system has a series
structure, connected by Division 1, Division 5, and a parallel subsystem, composed
of Divisions 2, 3, and 4. Based on this series-parallel expression of the system, the
system and division efficiencies have the following relationship:

E 0 = E 0 × ωð2Þ E 0 þ ωð3Þ E 0 þ ωð4Þ E0 × E0
where ωð2Þ = 6g = 1 wg Z g0 = 18

g = 1 wg Z g0 , ω
ð3Þ
= 12
g = 7 wg Z g0 =
18
g = 1 wg Z g0 , and
18 18
ω = g = 13 wg Z g0 = g = 1 wg Z g0 : Which of the first team offensive, bench team
(4)
offensive, or defensive divisions is more efficient is revealed by their efficiency

scores, and which of them has a greater impact on the performance of the system
depends on the value of their associated weight.
Series and parallel are the two basic structures of network systems. All network
structures that cannot be classified into these two types of structures are classified as
mixed structures and are discussed in this chapter. The discussion is based on the
number of divisions that constitute the network system. For three-division systems,
in addition to those discussed in Sect. 14.1, Ebrahimnejad et al. (2014) used a
division parameters distance function model to measure the performance of
49 branches of an Iranian bank with the same structure of Fig. 14.1, where the first
stage was composed of consumer and business banking to collect deposits for the
second stage to earn profit. Yu and Fan (2009) used the same model proposed in Yu
and Lin (2008) to measure the performance of 23 bus companies in Taiwan, with the
passenger and freight divisions in the production stage being replaced with highway
and urban ones. Amirteimoori and Yang (2014) measured the efficiency of 17 plants
in Iran that produced prefabricated cabins with the same structure of Fig. 14.2, where
different parts were manufactured in two divisions for the third division to assemble.
Adler et al. (2013) analyzed the performance of 43 airports in 13 European countries,
where two stages of operations were identified. The system has a structure the same
as that of Fig. 14.3, where the first stage had one division of generating passengers
and cargo, and the second had two divisions of aeronautical and non-aeronautical
activities. The two stages were evaluated independently, and the sum of the two
output division distance parameters was maximized for the two divisions in the
second stage. Lovell et al. (1994) investigated the performance of secondary educa-
tion in the US, where performance was defined as the ability of secondary schools to
convert human, physical, and financial resources into educational opportunities in
Stage one, which, together with student input, produced both intermediate- and long-
term educational outcomes in Stage two. The two divisions in Stage two also
consume exogenous inputs. A sample of 1032 high schools in the US in the base
year 1979–1980 was used to measure the efficiency of each division independently.
Yu (2010) constructed a slacks-based model to investigate the performance of
15 domestic airports in Taiwan by decomposing the operations into production
and services, where the latter were further separated into airside and landside. The
problem had the same structure as that in Lovell et al. (1994). The model showed that
the system efficiency of the airport was the product of the production and service
efficiencies, and the latter was a weighted harmonic average of airside and landside
service efficiencies. Amatatsu et al. (2012) evaluated industrial efficiency and the
corresponding returns to scale and discussed the proposed merger of 47 prefectures
by the Japanese government. The system is composed of three divisions, including
primary, secondary, and tertiary sectors. The input-output tables of regional econo-
mies were used as the data. The structure of the system is similar to that in Fig. 14.8,
except there are only three divisions in the matrix structure.
For four-division systems, Tavassoli et al. (2014) used the same structure as that
in Fig. 14.5 to investigate the performance of 11 domestic airlines in Iran from the
aspects of passengers and cargo using a slacks-based model. Both aspects have two
divisions, production and consumption. Prieto and Zofío (2007) undertook network
efficiency analysis within an input-output model that allowed them to assess poten-
tial technical efficiency gains by comparing technologies corresponding to different
economies. Different sectors use primary inputs to produce intermediate inputs and
outputs, and then satisfy the final demand. An output system distance measure model
was used to measure the performance of five OECD countries with four sectors
(divisions).
References
Adler, N., Liebert, V., & Yazhemsky, E. (2013). Benchmarking airports from a managerial
perspective. Omega, 41, 442–458.
Amatatsu, H., Ueda, T., & Amatatsu, Y. (2012). Efficiency and returns-to-scale of local govern-
ments. Journal of the Operational Research Society, 63, 299–305.
Amirteimoori, A., & Yang, F. (2014). A DEA model for two-stage parallel-series production
processes. Rairo-Operations Research, 48, 123–134.
Avkiran, N. K. (2009). Opening the black box of efficiency analysis: An illustration with UAE
banks. Omega, 37, 930–941.
References 393
Cheng, Y., & Gao, H. L. (2015). Matrix-type network DEA model with its application based on
input-output tables. Mathematical Problems in Engineering, 2015, 8. https://doi.org/10.1155/
2015/505941
Cook, W. D., Zhu, J., Bi, G. B., & Yang, F. (2010). Network DEA: Additive efficiency decompo-
sition. European Journal of Operational Research, 207, 1122–1129.
Ebrahimnejad, A., Tavana, M., Lotfi, F. H., Shahverdi, R., & Yousefpour, M. (2014). A three-stage
data envelopment analysis model with application to banking industry. Measurement, 49,
308–319.
Hsieh, L. F., & Lin, L. H. (2010). A performance evaluation model for international tourist hotels in
Taiwan: An application of the relational network DEA. International Journal of Hospitality
Huang, C. W., Chiu, Y. H., Fang, W. T., & Shen, N. (2014). Assessing the performance of Taiwan’s
environmental protection system with a non-radial network DEA. Energy Policy, 74, 547–556.
Lewis, H. F., & Sexton, T. R. (2004). Network DEA: Efficiency analysis of organizations with
complex internal structure. Computers and Operations Research, 31, 1365–1410.
Lin, T. Y., & Chiu, S. H. (2013). Using independent component analysis and network DEA to
improve bank performance evaluation. Economic Modelling, 32, 608–616.
Lovell, C. A. K., Walters, L. C., & Wood, L. L. (1994). Stratified models of education production
using modified DEA and regression analysis. In A. Charnes, W. W. Cooper, A. Y. Lewin, &
L. M. Seiford (Eds.), Data envelopment analysis: Theory, methodology and applications
(pp. 329–351). Kluwer.
Moreno, P., & Lozano, S. (2014). A network DEA assessment of team efficiency in the NBA.
Prieto, A. M., & Zofío, J. L. (2007). Network DEA efficiency in input-output models: With an
application to OECD countries. European Journal of Operational Research, 178, 292–304.
Tavassoli, M., Faramarzi, G. R., & Saen, R. F. (2014). Efficiency and effectiveness in airline
performance using a SBM-NDEA model in the presence of shared input. Journal of Air
Transport Management, 34, 146–153.
Wang, C. H., Lu, Y. H., Huang, C. W., & Lee, J. Y. (2013). R&D, productivity, and market value:
An empirical study from high-technology firms. Omega, 41, 143–155.
Yu, M. M. (2010). Assessment of airport performance using the SBM-NDEA model. Omega, 38,
440–452.
Yu, M. M., & Fan, C. K. (2009). Measuring the performance of multimode bus transit: A mixed
structure network DEA model. Transportation Research Part E-Logistics and Transportation
Review, 45, 501–515.
Yu, M. M., & Lin, E. T. J. (2008). Efficiency and effectiveness in railway performance using a
multi-activity network DEA model. Omega, 36, 1005–1017.
Zhao, Y., Triantis, K., Murray-Tuite, P., & Edara, P. (2011). Performance measurement of a
transportation network with a downtown space reservation system: A network-DEA approach.
Transportation Research Part E-Logistics and Transportation Review, 47, 1140–1159.
Chapter 15
Dynamic Systems
The efficiency measurements discussed thus far focus on static models, which either
assume that the inputs are fully used for producing outputs in the specified period of
time, or assume steady-state production, where capital or other quasi-fixed inputs are
fixed. There is thus no time interdependence between the input utilization and output
realization for a production unit in consecutive time periods, and these models do not
differentiate capital inputs from variable inputs. In reality, however, quasi-fixed
factors can change in the medium- and long-run, which introduces an intertemporal
effect with regard to inputs. In order to capture the effects caused by changes in the
quasi-fixed factors, a dynamic analysis is necessary.
The time interdependence among different periods can be ascribed to a number of
factors associated with the dynamic aspects of production, which are classified as
production delays, inventories, quasi-fixed inputs, adjustment costs, and incremental
improvements by Fallah-Fini et al. (2014). Production delays occur when inputs
contribute to both current and future output productions, and typical examples
include the effects of R&D activities on patents, advertising on sales, and human
resources on firm performance. Ignoring the lagged productive effects of such inputs
yields misleading efficiency measures. Inventories serve as buffers for leveling
resource consumption and fulfilling over-capacitated demand. Allowing for inven-
tories of exogenous inputs, intermediate products, and final outputs leads to
intertemporal inventory transfer in a production unit, and contributes directly to
the dynamic nature of production. Regarding capital or generally quasi-fixed inputs,
the values of these may depreciate, and if this is not taken into account then the
performance of some of the later periods of a production unit will be understated.
Moreover, the allocative and technical inefficiencies will also be confounded.
Adjustment costs for a production unit, typically in the form of foregone outputs
or reduced resources, are due to investment in new capital. For example, installing a
new machine usually results in delayed production, and thus a trade-off occurs
between current production and expansion for future production. Finally, firms
tend to learn and adjust production once they obtain the required information
https://doi.org/10.1007/978-3-031-27593-7_15
396 15 Dynamic Systems
about the sources of technical inefficiency. The presence of all these factors makes a
static model insufficient to characterize the dynamic nature of the production
process. Dynamic models are thus needed to provide insights on the behavior of
the overall production system.
This chapter discusses efficiency measurement in dynamic environments. The
models are classified as ratio-form, distance function, slacks-based, and value-based
ones, and are discussed in Sects. 15.1 to 15.4, respectively. In addition to the basic
dynamic structure, where each period is represented by a whole-unit (black-box)
system, dynamic network structures, where each period is represented by a network
system, are also presented. Finally, in the Supplementary Literature section different
dynamic systems that have appeared in the literature are briefly described.
ðt Þ
The basic dynamic system is a sequence of periods linked by carryovers K l , as
ðt Þ
depicted in Fig. 15.1, where the inputs X i supplied from outside in each period t,
ðt - 1Þ
together with the carryovers K l transferred from the preceding period t-1,
ðt Þ ðt Þ
produce the outputs Y r that are sent out of the system, and the carryovers K l for
the succeeding period t + 1 to use for production.
Structurally, the basic dynamic system resembles the multi-stage system (refer-
ðt Þ
ring to Fig. 12.1). The difference is that the inputs X i , outputs Y ðrtÞ , and carryovers
ðt Þ
K l are the same for every period. The basic dynamic system is thus a special case of
a multi-stage system, such that the properties possessed by the latter are also
possessed by the former.
X i(1) , i = 1, ..., m X i(t ) X i( p ) , i = 1, ..., m
K l( 0) K l(1) K l(t –1) K l(t ) K l( p–1) K l( p )

… t … p
l=1,..., q l=1,..., q
Yr(1) , r = 1, ..., s Yr(t ) Yr( p ) , r = 1, ..., s
Fig. 15.1 The basic dynamic structure

15.1.1 The Whole-unit Case
Based on the basic dynamic structure shown in Fig. 15.1, Kao (2013) measured the
performance of eight forest districts in Taiwan, where the forest stock plays the role
of carryover. If more forest stock is harvested in a period, then the timber production
of this period, as an output, will be high. On the other hand, the forest stock, as a
carryover left for succeeding periods to harvest, becomes low. The forest stock must
thus be taken into consideration in measuring efficiencies, or the period-specific
efficiencies calculated from the conventional static model will be misleading.
For a planning horizon of p periods, Kao (2013) proposed the following output-
oriented model under constant returns to scale to measure the efficiency of a forest
district:
p m ðtÞ q ð0Þ
1 t=1 i = 1 vi X i0 þ l = 1 cl K l0
= min: p s ðtÞ q ðpÞ
r = 1 ur Y r0 þ
E0
t=1 l = 1 cl K l0
m q s q
s:t:
ðtÞ
vi X i j þ
ðt - 1Þ
cl K l j -
ðtÞ
ur Y r j þ
ðtÞ
cl K l j ≥ 0, ð15:1Þ
i=1 l=1 r=1 l=1
j = 1, . . . ,n, t = 1, . . . , p
ur ,vi ,cl ≥ ε, r = 1, . . . ,s, i = 1, . . . ,m, l = 1, . . . , q
variable substitution technique introduced in Charnes and Cooper (1962).
Model (15.1) is a relational one, in that the same multiplier is attached to the same
factor, which reduces the number of variables in a scale of p. Once an optimal
solution (u*, v*, c*) is obtained, the efficiencies of the forest district for the whole
ðt Þ
p periods (E0) and period t (E 0 ) are calculated as:
p s ðt Þ q ðpÞ
t=1 r = 1 ur Y r0 þ l = 1 cl K l0
E0 = p m ðt Þ q ð0Þ
t=1 i = 1 vi X i0 þ l = 1 cl K l0
ð15:2Þ
s ðt Þ q ðt Þ
ðt Þ r = 1 ur Y r0 þ l = 1 cl K l0
E0 = , t = 1, . . . , p
m ðt Þ q ðt - 1Þ
i = 1 vi X i0 þ l = 1 cl K l0
As discussed in Chap. 12, the efficiency for the whole p periods is greater than or
equal to the product of those of p periods and is less than or equal to the weighted
ðt Þ ðt Þ
average of those of the p periods, i.e., pt = 1 E 0 ≤ E 0 ≤ pt = 1 ωðtÞ E0 , where the
weight ω associated with period t is the proportion of the aggregate input con-
(t)
sumed in this period in that consumed in all p periods.

ðt Þ
If the carryovers are ignored, then the efficiency of period t is sr = 1 ur Y r0 =
m ðt Þ ðt Þ
i = 1 vi X i0 . Compared to E 0 in Eq. (15.2), it is noted that the efficiency of the
ðt - 1Þ
periods receiving a larger amount of K l from the preceding period will be
ðt Þ
overstated, while it will be understated for the periods leaving a larger amount of K l
to the succeeding period.
15.1.2 The Network Case
With regard to measuring the performance of a production unit over time when
different strategies are applied to achieve this, Akbarian et al. (2015) proposed
constructing a network structure showing the cause-effect relationships of the key
performance indicators (KPIs) obtained from a balanced score card (BSC), and used
the nine oil refineries of the National Iranian Oil Refining and Distribution Company
as an example to illustrate this approach. Figure 15.2 shows the structure of the
network. There are two processes, with four divisions in each, learning, internal,
ðt Þ
customer, and financial. The KPIs are training (Z 1 ), gas production in the refinery
ðt Þ ðt Þ ðt Þ
(Z 2 ), material assigned to refinery (Z 3 ), quality of oil products (Z 4 ), and
ðt Þ
management systems (Z 5 ), indicating the cause-effect relationships for divisions
of the same process, and customer portal changes (K1), effective measuring of
training (K2), increases in refinery capacity (K3), strategic plans (K4), outsourcing
t1 process t 2 process
K1
X Learning Learning
K2
Z1(1) Z1(2)
K3
Internal Internal
Z (1) Z 3(2)
3 K5 Z 2( 2)
Z 2(1)
Customer Customer
K4
Z 4(1) Z 4( 2)
K6
Financial Financial
Z 5(1) Z 5( 2)
Fig. 15.2 Structure of the oil refinery example

(K5), and safety stock for refinery (K6), indicating the cause-effect relationships for
divisions of two consecutive processes, with annual budget (X) as the input and net
profit margin (Y ) as the final output.
ð1Þ
To measure the efficiencies of the whole system (E0), the two processes (E 0 and
ð2Þ ðt Þ½k
E 0 ), and the four divisions in the two processes (E 0 , t = 1, 2, k = 1, . . . , 4 ),
Akbarian et al. (2015) assumed the same factors in different processes were differ-
ent, and thus assigned different multipliers to them. Based on the structure shown in
Fig. 15.2, the efficiencies are calculated as follows:
E0 = uY 0 =vX 0
ð1Þ 6
E0 = c K =vX 0
l = 1 l l0
ð2Þ 6
E0 = uY 0 = l = 1 cl K l0
ð1Þ½1 ð1Þ ð1Þ
E0 = w1 Z 10 þ c1 K 10 þ c2 K 20 =vX 0
ð1Þ½2 ð1Þ ð1Þ ð1Þ ð1Þ ð1Þ ð1Þ ð1Þ ð1Þ
E0 = w2 Z 20 þ w3 Z 30 þ c3 K 30 þ c4 K 40 þ c5 K 50 = w1 Z 10 þ w5 Z 50
ð1Þ½3 ð1Þ ð1Þ ð1Þ ð1Þ
E0 = w4 Z 40 =w2 Z 20
ð1Þ½4 ð1Þ ð1Þ ð1Þ ð1Þ ð1Þ ð1Þ
E0 = w5 Z 50 þ c6 K 60 = w3 Z 30 þ w4 Z 40
E0 = w1 Z 10 =c6 K 60
ð2Þ½2 ð2Þ ð2Þ ð2Þ ð2Þ ð2Þ ð2Þ ð2Þ ð2Þ
E0 = w2 Z 20 þ w3 Z 30 = w1 Z 10 þ w5 Z 50 þ c3 K 30
ð2Þ½3 ð2Þ ð2Þ ð2Þ ð2Þ
E0 = w4 Z 40 = w2 Z 20 þ c1 K 10 þ c4 K 40
ð2Þ½4 ð2Þ ð2Þ ð2Þ ð2Þ ð2Þ ð2Þ
E0 = uY 0 þ w5 Z 50 = w3 Z 30 þ w4 Z 40 þ c2 K 20 þ c5 K 50
ð15:3Þ
The efficiency of the whole system is the product of those of the two processes,
ð1Þ ð2Þ
E0 = E0 × E0 :
To obtain the optimal multipliers to calculate these efficiencies, Akbarian et al.
(2015) proposed an efficiency aggregation model, with the objective function being
a weighted average of the four divisions in the two processes, where the weight
associated with a division is the proportion of the aggregate input consumed by this
division in that consumed by all divisions in the two periods, under the constraint
that all the efficiencies calculated from Eq. (15.3) must be less than or equal to one.
In symbols, the model is:
2 5 ðt Þ ðt Þ 6
uY 0 þ t=1 g = 1 wg Z g0 þ l = 1 cl K l0
max: 2 5 ðt Þ ðt Þ 6
vX 0 þ t=1 g = 1 wg Z g0 þ l = 1 cl K l0
ðt Þ½k
s:t: Ej ≤ 1, t = 1, 2, k = 1, . . . , 4, j = 1, . . . , n
ð15:4Þ
ðt Þ
E j ≤ 1, t = 1, 2, j = 1, . . . , n
Ej ≤ 1, j = 1, . . . , n
u, v, wðgtÞ , cl ≥ ε, t = 1, 2, g = 1, . . . , 5, l = 1, . . . , 6
Note that the objective function is used just as a direction for finding a set of
optimal multipliers to calculate the efficiencies defined in Eq. (15.3). It is not the
efficiency of the whole system.
Early studies on measuring the efficiency of dynamic systems were based on the
network DEA model developed by Färe and Grosskopf (1996, 2000). The output-
oriented model for the basic dynamic system shown in Fig. 15.1 under constant
returns to scale is:
max: φ
n
ðtÞ ðtÞ ðtÞ
s:t: λj X i j ≤ X i0 , i = 1, . . . , m, t = 1, ..., p
j=1
n
ðtÞ ðtÞ ðtÞ
λj Y i j ≥ φY r0 , r = 1, . . . , s, t = 1, ..., p
j=1
n
ðtÞ ðt - 1Þ ðt - 1Þ
λj K l j ≤ K l0 , l = 1, . . . , q, t = 1, ..., p ð15:5Þ
j=1
n
ðtÞ ðtÞ ðtÞ
λj K l j ≥ K l0 , l = 1, . . . , q, t = 1, ..., p-1
j=1
n
ðpÞ ðpÞ ðpÞ
λj K l j ≥ φK l0 , l = 1, ..., q
j=1
ðtÞ
λj ≥ 0, j = 1, . . . , n, t = 1, ..., p
Here the non-Archimedean number ε is ignored for simplicity of expression. The

ðt Þ
convexity constraint of nj = 1 λj = 1 can be added if the production technology of
period t is assumed to be variable returns to scale.
p m p q p-1 q
1 ðt Þ ðt Þ ðt Þ ðt - 1Þ ðt Þ ðt Þ
= min: vi X i0 þ cl K l0 - cl K l0
E0 t=1 i=1 t=1 l=1 t=1 l=1
p s q
ðt Þ ðpÞ ðpÞ
s:t: uðrtÞ Y r0 þ cl K l0 =1
t=1 r=1 l=1
m q s q
ðt Þ ðt Þ ðt Þ ðt - 1Þ ðt Þ ðt Þ ðt Þ
vi X ij þ cl K lj - uðrtÞ Y rj þ cl K lj ≥ 0,
i=1 l=1 r=1 l=1
j = 1, . . . , n, t = 1, . . . , p
uðrtÞ , vi , cl , cl ≥ 0, 8r, i, l, t
ð15:6Þ
Suppose the multipliers associated with the same factor of different periods are
required to be the same, i.e., uðrtÞ = ur , vi = vi , cl = cl = cl : Model (15.6) boils
down to the linearized version of Model (15.1). Note that the superscript t in the last
term of the objective function runs from 1 to p-1, rather than p. In other words, the
ratio-form model proposed by Kao (2013) is a special case of the general distance
function model developed by Färe and Grosskopf (1996, 2000) when the non-
Archimedean number ε is ignored. The objective value of the former is thus greater
than or equal to that of the latter, which implies that the efficiency measured from
Model (15.1) is less than or equal to that measured from Model (15.6). From this
aspect, the model proposed by Kao (2013) is effective in identifying the order of
performance for a set of DMUs.
15.2.1 The Production Delays Example
Dynamic models use carryovers to describe the interdependence of two consecutive

periods, and there are different ways to express carryovers. Extending the idea of
Chen (2009), Chen and van Dalen (2010) proposed a distance function model to
measure the efficiency of a system with lagged productive effects, where the inputs
of a period contribute to both current and future output production. The lagged
effects of R&D activities and advertising are typical examples.
Figure 15.3 is a pictorial illustration of production delays. In the conventional
production with no delays, the inputs of a period t produce the outputs in the same
period t. When production delays occur, the inputs of period t contribute to products
of not only the current period, expressed by a solid arrow but also future q periods,
expressed by dashed arrows. This also implies that outputs of period t are attributed
to not only the inputs of the current period but also those of q preceding ones. Under
stationary conditions the effect of the inputs at period t on the outputs of period t + l,
l = 1, . . ., q, is the same, and is assumed to be ω(l ), with ql = 0 ωðlÞ = 1 and ω(l ) ≥ 0.
ðt Þ
In other words, a DMU applies the inputs X i0 in period t to produce the total outputs
Fig. 15.3 Pictorial X i(t –1) X i(t ) X i(t +1) X i(t + q ) X i(t +q +1)
illustration of production
delays
… … … … …
t–1 t t+1 … t+q t+q+1
Yr(t –1) Yr(t ) Yr( t +1) Yr(t + q ) Yr(t + q +1)
ðt Þ ðtþlÞ
of Y r0 = ql = 0 ωðlÞ Y r0 , r = 1, . . ., s, in q + 1 different periods. The output-oriented
model under constant returns to scale is:
max:φðtÞ
n
ðtÞ ðtÞ
s:t: λj X i j ≤ X i0 , i = 1, . . . , m
j=1
n ð15:7Þ
ðtÞ ðt Þ ðt Þ
λj Y~ rj ≥ φðtÞ Y~ r0 , r = 1, . . . , s
j=1
ðtÞ
λj ≥ 0, j = 1, . . . , n
The efficiency of a DMU can be different in different periods. This model can be
extended to cover more than q + 1 periods to measure the overall efficiency over a
desired horizon.
Model (15.7) was used by Chen and van Dalen (2010) to evaluate the advertising
efficiencies of seven automobile producers and eight pharmaceutical companies in
North America, with panel data ranging from 1997 to 2005.
15.2.2 The Period Distance Parameters Case
The characteristic of a dynamic system is that some outputs of one period can be
carried over to the next as finished goods inventory, and some of the inputs can be
stored and used in future periods. Let Y Fr ðtÞ and Y Kr ðtÞ denote the finished output and
Sð t Þ
that to be carried over to the next period, respectively, and X i denote the input
stored for future use, in period t. An output model for measuring the efficiency of the
system proposed by Färe and Grosskopf (1996), under constant returns to scale and
based on the structure shown in Fig. 15.4, is as follows:
Fig. 15.4 Dynamic systems X iS (t – 2) + X iS (t –1) + X iS (t ) +

with storable inputs and
X i(t –1) – X i(t ) – X i(t + 1) –
carryover outputs
X iS (t –1) X iS (t ) X iS (t + 1)
YrK (t –2) YrK (t –1) YrK (t ) YrK (t +1)

t–1 t t+1
YrF (t –1) YrF (t ) YrF (t +1)
p
max: φðtÞ
t=1
n
ðtÞ Sðt - 1Þ ðtÞ SðtÞ Sðt - 1Þ ðtÞ SðtÞ
s:t: λj Xi j þ Xi j ; - Xi j ≤ X i0 þ X i0 - X i0 ,
j=1
i = 1, . . . ,m,t = 1, . . . , p
n ð15:8Þ
ðtÞ
≥ φðtÞ Y r0 þ Y r0 ,
FðtÞ KðtÞ FðtÞ KðtÞ
λj Yr j þ Yr j r = 1, . . . ,s, t = 1, . . . , p
j=1
n
ðtÞ Kðt - 1Þ Kðt - 1Þ
λj Y r j ≤ Y r0 , r = 1, . . . ,s, t = 1, . . . , p
j=1
ðtÞ
φðtÞ ≥ 1,λj ≥ 0, j = 1, . . . ,n, t = 1, . . . , p
In this model each period is allowed to have different distance parameters φ(t),
which provide the possibility of changes in technology throughout the planning
horizon. The efficiency of the system can be defined as the average of the p φ(t) ’s
(in reciprocal form).
15.2.3 Directional Distance Function: Whole Unit
Directional distance function models are widely used for identifying inefficiencies
when there are undesirable factors. For example, Skevas et al. (2012) investigated
the performance of Dutch arable farms by using a Russell type measure to identify
technical and pesticides’ environmental inefficiencies.
V ðt Þ
As depicted in Fig. 15.5, a farm uses variable inputs X i , i = 1, . . ., m1, and fixed
F ðt Þ ðt Þ
inputs X i , i = m1 + 1, . . ., m, to produce outputs Y r in period t. The production of
ðt Þ
Y ðrtÞ involves an indirect effect related to pesticide use El (i.e., environment impact).
Pesticides impact the production process in the current period by decreasing pest
F (t )
Fig. 15.5 A dynamic X iV (t ) X i
system with undesirable
factors
El(t –1) El(t )

t
Yr(t ) El(t )
damage and in the next period through their negative impact on organisms that can
ðt Þ
benefit the farm. Assuming weak disposability of the undesirable factors E l , and
F ðt Þ
fixed inputs X i , the model proposed by Skevas et al. (2012) is as follows:
ðtÞ ðtÞ ðtÞ ðtÞ

max:η1 þ η2 þ η3 þ η4
n
ðtÞ VðtÞ VðtÞ ðtÞ ðtÞ
s:t: λj X i j ≤ X i0 - η1 f i , i = 1, . . . , m1
j=1
n
ðtÞ FðtÞ FðtÞ
λj X i j = σX i0 , i = m1 þ 1, . . . , m
j=1
n
ðtÞ ðtÞ ðtÞ ðtÞ
λj Y r j ≥ Y r0 þ η2 gðtÞ
r , r = 1, . . . , s ð15:9Þ
j=1
n
ðtÞ ðtÞ ðtÞ ðtÞ ðtÞ
λj El j = σ El0 - η3 hl , l = 1, . . . , q
j=1
n
ðtÞ ðt - 1Þ ðt - 1Þ ðtÞ ðtÞ
λj E l j = σ E l0 - η4 h l , l = 1, . . . , q
j=1
ðtÞ
0 ≤ σ ≤ 1, λj ≥ 0, j = 1, . . . , n
where (f, g, h) is a pre-specified vector of direction. It seeks to increase the desirable

outputs while simultaneously reducing the undesirable outputs and variable inputs.
ðt Þ ðt Þ ðt Þ ðt Þ
Specifically, η1 , η2 , η3 , and η4 correspond to the free disposability of variable
V ðt Þ ðt Þ
inputs X i and desirable outputs Y ðrtÞ , undesirable outputs E l and undesirable
ðt - 1Þ
inputs E l , respectively. The scaling parameter σ is selected to ensure a feasible
solution with weakly disposable fixed inputs and undesirable factors under variable
returns to scale. A value of one indicates that the optimal inputs or outputs can be
freely achieved (i.e., there is strong disposability), and a value strictly less than one
indicates that they cannot be achieved freely (i.e., there is weak disposability).
This model is able to decompose the technical inefficiency of the different inputs
and outputs for each firm. As an illustration, 703 observations from 188 farms in the
Netherlands over the period 2003–2007 were collected, and the efficiencies
measured.
15.2.4 Directional Distance Function: Network
The production system in each period can also have a network structure. Chen
(2012) studied pig production in Taiwan, in which the entire production was divided
into two processes, breed-to-farrow and wean-to-finish, with the structure shown in
Fig. 15.6. The activities in the breed-to-farrow process are breeding females and their
maintenance during pregnancy and nursing, and those in the wean-to-finish process
are the care and feeding of pigs after weaning and until they reach a butchery weight.
ðt Þ
The inputs for the breed-to-farrow phase are boars and sows (X i ), and for the wean-
ðt Þ
to-finish phase are the weaned pigs (K l ). There are also inputs shared by the two
Sð t Þ ðt Þ
processes, such as labor, capital, and feed (X i ). The outputs are weaned pigs (K l )
for the breed-to-farrow process, which also play the role of carryover, and finished
hogs (Y ðrtÞ ) for the wean-to-finish process.
Suppose the shared inputs are used by the two processes with unknown fractions
of α and 1 - α. Chen (2012) proposed the following directional distance function
model to measure the efficiency of this system:
Fig. 15.6 Structure of the X i(t ) X i(t +1)

pig production system
t t+1
α Breed-to- K l(t ) α Breed-to- K l(t +1)

farrow farrow
X iS (t ) X iS (t +1)
1– α 1– α
K l(t –1) Wean-to- Wean-to-
finish finish
Yr(t ) Yr(t +1)

p
ðtÞ ðtÞ
max: ξðtÞ ωð1Þ η1 þ ωð2Þ η2
t =1
n
ðtÞ½1 ðtÞ ðtÞ ðtÞ
s:t: λj X i j ≤ 1 - η1 X i0 , i = 1, ...,m1 , t = 1, . .., p
j =1
n
ðtÞ½1 ðtÞ SðtÞ ðtÞ ðtÞ SðtÞ
λj αi X i j ≤ 1- η1 αi X i0 , i= m1 þ 1,. ..,m, t = 1, ..., p
j =1
n
ðtÞ½2 ðtÞ SðtÞ ðtÞ ðtÞ SðtÞ
λj 1 - αi X i j ≤ 1 - η1 1 - αi X i0 , i = m1 þ 1,.. .,m, t = 1, . . ., p
j =1
n
ðtÞ½2 ðtÞ ðtÞ ðtÞ
λj Yr j ≥ 1 þ η2 Y r0 , r = 1, ... ,s, t = 1, .. .,p
j =1
n n
ðtÞ½1 ðtÞ ðtþ1Þ½2 ðtÞ
λj Kl j ≥ λj Kl j , l = 1, ...,q, t = 1, .. .,p
j =1 j=1
n
λj K l j ≤ K l0 , l = 1, .. ., q
j =1
ðtÞ ðtÞ ðtÞ
Li ≤ αi ≤ U i , i = m1 þ 1, .. .,m, t = 1, . . ., p
ðtÞ
λj ≥ 0, j = 1, ...,n, t = 1, ... ,p
ð15:10Þ
where ξ(t), ω(1), and ω(2) are pre-specified weights, indicating the relative importance
of period t, breed-to-farrow, and wean-to-finish processes, with pt = 1 ξðtÞ = 1 and
ω(1) + ω(2) = 1. Numbers in parentheses represent periods and those in square
ðt Þ ðt Þ
brackets represent divisions. Li and U i are the lower and upper bounds, respec-
tively, for the unknown fraction αi . Positive values of η1 and η2 indicate that the
corresponding process in period t is inefficient. The performances of 29 pig farms in
Taiwan were evaluated using this model.
Tone (2001) proposed the slacks-based approach to measure the efficiency of the
whole-unit (black box) system, which is equivalent to the Russell ratio efficiency,
and Tone and Tsutsui (2009) extended the idea to network systems. Structurally, a
basic dynamic system resembles a whole-unit system repeated for a number of
periods. The model developed for network systems can thus also be applied to
dynamic systems, as discussed in Tone and Tsutsui (2010, 2014). Their model has
three forms, input-, output-, and non-oriented (see Chap. 4), where the last one is a
combination of the first two, and is thus termed combined model in this chapter. In
Subsection 15.3.1 we will introduce the combined form, and then discuss two ideas
for the modeling of this approach in Subsections 15.3.2 and 15.3.3.
15.3.1 The Basic Dynamic Structure Case
For the basic dynamic structure shown in Fig. 15.1, the combined slacks-based
model for measuring efficiency is as follows:
p ðt Þ 1 m ðt Þ - ðt Þ q ðt - 1Þ - ðt - 1Þ
t = 1ω 1-
mþq
s
i=1 i
=X i0 þ s
l=1 l
=K l0
min:
1 s ðt Þ q ðt Þþ ðt Þ
p
t = 1ω
ðt Þ 1þ sðtÞþ =Y r0 þ s =K l0
sþq r=1 r l=1 l
n
ðt Þ ðt Þ ðt Þ - ðt Þ
s:t: λj X ij þ si = X i0 , i = 1, . . . , m, t = 1, . . . , p
j=1
n
λj Y rj - sðrtÞþ = Y r0 , r = 1, . . . , s, t = 1, . . . , p
j=1
n
ðt Þ ðt - 1Þ ðt - 1Þ - ðt - 1Þ
λj K lj þ sl = K l0 , l = 1, . . . , q, t = 1, . . . , p
j=1
n
ðt Þ ðt Þ ðt Þþ ðt Þ
λj K lj - sl = K l0 , l = 1, . . . , q, t = 1, . . . , p
j=1
n n
ðt Þ ðt Þ ðtþ1Þ ðtþ1Þ
λj K lj = λj K lj , l = 1, . . . , q - 1
j=1 j=1
λ, s ≥ 0
ð15:11Þ
where ω(t) in the objective function are pre-specified weights, representing the
relative importance of period t in measuring the aggregate efficiency of all
p periods, which are required to be nonnegative, with pt = 1 ωðtÞ = 1: The constraints
can be modified appropriately for non-discretionary and undesirable factors. Under
ðt Þ
variable returns to scale the constraint of nj = 1 λj = 1 is added. This model is a
linear fractional program, which can be linearized by applying a variable substitution
technique, as described in Chap. 4.
The objective function of this model expresses the system efficiency as a combi-
nation of the weighted average of the period-specific input efficiencies and that of the
period-specific output efficiencies. Since the efficiency of each period t is:
m ðt Þ - ðt Þ q ð t - 1Þ - ðt - 1Þ
1- 1
i = 1 si =X i0 þ l = 1 sl =K l0
ðt Þ mþq
E0 = , t = 1, . . . , p ð15:12Þ
s ðt Þþ ðt Þ q ðt Þþ ðt Þ
1 þ sþq
1
r = 1 sr =Y r0 þ l = 1 sl =K l0
A more reasonable expression for the system efficiency is a weighted average of the
ðt Þ
period efficiencies, E 0 = pt = 1 ωðtÞ E0 , which is nonlinear. To maintain linearity,
Kao (2014) proposed defining the weight ω(t) of period t as the ratio of the output
efficiency (in reciprocal form) of this period to that of all p periods, that is,
s ðt Þþ ðt Þ q ðt Þþ ðt Þ
1 þ sþq
1
r = 1 sr =Y r0 þ l = 1 sl =K l0
ðt Þ
ω = , t = 1, . . . , p ð15:13Þ
p s ðt Þþ ðt Þ q ðt Þþ ðt Þ
t=1 1 þ sþq
1
r = 1 sr =Y r0 þ l = 1 sl =K l0
In this case the weighted average of the efficiencies in Eq. (15.11) and the weights
in Eq. (15.13) becomes:
p m ðt Þ - ðt Þ q ðt - 1Þ - ð t - 1Þ
t=1 1- 1
mþq i = 1 si =X i0 þ l = 1 sl =K l0
E0 = ð15:14Þ
p s ðt Þþ ðt Þ q ðt Þþ ðt Þ
t=1 1 þ sþq
1
r = 1 sr =Y r0 þ l = 1 sl =K l0
The associated model is to minimize E0, as defined in Eq. (15.14), subject to the
constraints of Model (15.11).
Different from Model (15.10), where the weights are pre-specified subjectively,
the weights of the model proposed by Kao (2014) are determined objectively from
the data, which are the most favorable ones to yield the largest system efficiency.
15.3.2 The Aggregate Slack Case
Von Geymueller (2009) proposed a dynamic model to measure the efficiency of

50 US electricity transmission system operators in the period 2000–2006. The
factors considered are two inputs, transmission materials (X1) and transmission
salaries and wages (X2); two quasi-fixed inputs, transmission lines (K1) and total
installed transmission capacity (K2); and one output, transmission of electricity (Y ).
Instead of using the aggregate efficiency of all factors as the objective function, this
model uses the aggregate slack of all factors adjusted over time. The model, in
general form, is as follows:
p m s q q
ðt Þ - ðt - 1Þ - ðt Þþ
max: γt si þ sðrtÞþ þ sl þ sl
t=1 i=1 r=1 l=1 l=1
n
ðt Þ ðt Þ ðt Þ - ðt Þ
s:t: λj X ij þ si = X i0 , i = 1, . . . , m, t = 1, . . . , p
j=1
n
λj Y rj - sðrtÞþ = Y r0 , r = 1, . . . , s, t = 1, . . . , p
j=1 ð15:15Þ
n
ðt Þ ðt - 1Þ ðt - 1Þ - ðt - 1Þ
λj K lj þ sl = K l0 , l = 1, . . . , q, t = 1, . . . , p
j=1
n
ðt Þ ðt Þ ðt Þþ ðt Þ
λj K lj - sl = K l0 , l = 1, . . . , q, t = 1, . . . , p
j=1
λ, s ≥ 0
where γ is a discounting factor. At optimality, the efficiency of each period is

calculated as:
m ðt Þ - q ðt - 1Þ - s ðt Þ
ðt Þ 1 si sl Y r0
E0 ¼ 1- þ 1- þ
m þ 2q þ s i¼1 X i0
ðt Þ
l¼1 K l0
ðt - 1Þ
r¼1
ðt Þ
Y r0 þ sr
ðt Þþ
q ðt Þ
K l0
þ ðt Þ ðt Þþ
, t ¼ 1, . . . , p
l¼1 K l0 þ sl
and the efficiency of the system is the average of the period efficiencies:
p
1 ðt Þ
E0 = E0
p t=1
If the efficiency of the system is measured in a static manner, then the quasi-fixed
inputs, as the outputs in each period, are omitted, resulting in the objective
function of:
p m s q
ðt Þ - ðt - 1Þ -
max: γt si þ sðrtÞþ þ sl ð15:16Þ
t=1 i=1 r=1 l=1
Using this approach, von Geymueller (2009) found that the static model under-
states the efficiency of all 50 operators over the period 2000–2006.
15.3.3 The Network Case
Avkiran (2015) used an SBM model to measure the performance of 16 foreign and
32 domestic banks in China in the period 2008–2010. A bank network is conceptu-
alized as comprised of two divisions, interest-bearing and non-interest operations,
linked by the number of referrals (Z ). Undesirable outputs from the two divisions,
non-performing loans (K1) and proportion of fruitless referrals (K2), are treated as
carryovers that affect the efficiency of the following periods. The structure of the
½1
system is depicted in Fig. 15.7, where the inputs are personnel expenses (X 1 ), other
½1 ½1
operating expenses (X 2 ), interest expenses on customer deposits (X 3 ), and other
½1
interest expenses (X 4 ) for the interest-bearing operations (Division 1), and person-
½2 ½2
nel expenses (X 1 ) and other operating expenses (X 2 ) for the non-interest operations
(Division 2). The outputs are interest income and loans (Y1) and other interest
income (Y2) for Division 1, and net fees and commissions (Y3) and other operating
income (Y4) for Division 2.
Following the idea in Tone and Tsutsui (2014), the model for measuring the
efficiency of this dynamic system is as follows:
Fig. 15.7 Structure of the X 1(t )[1] ~ Y1( t ) ,

commercial banking X 4(t )[1] Y2( t )
example
t t+1
K1(t –1) Interest- K1( t ) Interest- K1(t +1)

bearing bearing
K 2(t –1) Non- In- K 2( t ) Non- In- K 2(t +1)

terest terest
X 1(t )[2] , Y3( t ) ,

X 2(t )[2] Y4( t )
ðt Þ½1 - ðt Þ½2 - ðt Þþ
p ðt Þ 1 4 si 2 si sðtÞ - 2 sl0
t = 1ω 1-
9 i=1 ðt Þ½1
þ i=1 ðt Þ½2
þ ðt Þ
þ l=1 ðt Þ
X i0 X i0 Z0 K l0
min: ðt - 1Þ -
p ðt Þ 1 4 sðrtÞþ sðtÞþ 2 sl0
t = 1ω 1þ
7 r=1 ðt Þ
þ ðt Þ
þ l=1 ðt - 1Þ
Y r0 Z0 K l0
n
ðt Þ½1 ðt Þ½1 ðt Þ½1 - ðt Þ½1
s:t: λj X ij þ si = X i0 , i = 1, . . . , 4, t = 1, . . . , p
j=1
n
ðt Þ½2 ðt Þ½2 ðt Þ½2 - ðt Þ½2
λj X ij þ si = X i0 , i = 1, 2, t = 1, . . . , p
j=1
n
ðt Þ½1 ðt Þ ðt Þ
λj Y rj - sðrtÞþ = Y r0 , r = 1, 2, t = 1, . . . , p
j=1
n
λj Y rj - sðrtÞþ = Y r0 , r = 3, 4, t = 1, . . . , p
j=1
n
λj Zj - sðtÞþ = Z 0 , t = 1, . . . , p
j=1
n
λj Zj þ sðtÞ - = Z 0 , t = 1, . . . , p
j=1
n
ðt Þ½1 ðt - 1Þ ðt - 1Þ - ðt - 1Þ
λj K 1j - s1 = K 10 , t = 1, . . . , p
j=1
n
ðt Þ½2 ðt - 1Þ ðt - 1Þ - ðt - 1Þ
λj K 2j - s2 = K 20 , t = 1, . . . , p
j=1
n
ðt Þ½1 ðt Þ ðt Þþ ðt Þ
λj K 1j þ s1 = K 10 , t = 1, . . . , p
j=1
n
ðt Þ½2 ðt Þ ðt Þþ ðt Þ
λj K 2j þ s2 = K 20 , t = 1, . . . , p
j=1
λ, s ≥ 0
ð15:17Þ
Note that the carryovers K1 and K2 are undesirable factors, the corresponding
constraints are thus modified appropriately, and the position of the slack variables in
the objective function are changed accordingly. The results are illustrative in nature
because of the simulated data used for two of the variables.
15.4 Value-Based Efficiency Measures
The value-based model is used to find the maximum of either the aggregate output
p s ðt Þ p s ðt Þ p m ðt Þ
t=1 r = 1 ur Y r0 or the total profit t=1 r = 1 Pr Y r0 - t=1 i = 1 C i X i0 , where
Pr and Ci are unit prices of the outputs and unit costs of the inputs, respectively. This
type of model has been used to identify the optimal path of adjustment for the input
quantities and provide a measure of the potential cost savings.
The classic work related to this type of model is Nemoto and Goto (1999), in
which the objective is to minimize the total cost incurred in the periods being
considered. Based on the basic dynamic structure shown in Fig. 15.1, the
corresponding value-based model is as follows:
p m q
ðtÞ ðtÞ ðtÞ ðt - 1Þ
min: γt W i xi þ V l kl
t=1 i=1 l=1
n
ðtÞ ðtÞ ðtÞ
s:t: λj X i j ≤ x i , i = 1, . . . ,m, t = 1, . . . , p
j=1
n
ðtÞ ðtÞ ðtÞ
λj Y r j ≥ Y r0 , r = 1, . . . ,s, t = 1, . . . , p
j=1
n
ðtÞ ðt - 1Þ ðt - 1Þ
λj K l j ≤ kl , l = 1, . . . ,q, t = 1, . . . , p
j=1
n
ðtÞ ðtÞ ðtÞ
λj K l j ≥ K l0 , l = 1, . . . ,q, t = 1, . . . , p
j=1
ð0Þ
k l0 = kl0 ,
ð0Þ
l = 1, . . . , q
ðtÞ ðtÞ ðt - 1Þ
λj ,xi ,k l ≥ 0, j = 1, . . . ,n, i = 1, . . . ,m, l = 1, . . . ,q, t = 1, . . . , p
ð15:18Þ
ðt Þ ðt Þ
where γ is a discounting factor, W i and V l are the values of the input Xi and
carryover Kl at time t, respectively. The initial amounts of the quasi-fixed inputs are
ð0Þ
given at kl0 . This model identifies the optimal paths of the variable inputs Xi and
quasi-fixed inputs Kl over the planning horizon p.
Treating the quasi-fixed factors as outputs in period t implies that the DMU needs
to give up a certain amount of its regular outputs to maintain more quasi-fixed inputs,
as is expressed by the constraint of nj = 1 λj K lj ≥ K l0 : However, maintaining more
quasi-fixed factors transfers the current production to future periods, as indicated by
ðt Þ ðt - 1Þ ðt - 1Þ
the constraint of nj = 1 λj K lj ≤ kl :
Different models can be applied to measure the efficiency of a dynamic system,
and which one to use is dependent on the objectives and assumptions of the study.
Dynamic systems are repetitions of the single-period systems connected by carry-

overs, where the single-period system can have any of the structures discussed in the
preceding chapters. Färe and Grosskopf (1996) is an early work on this topic. Fallah-
Fini et al. (2014) provided a review of the literature on dynamic efficiency measure-
ment models, and Mariz et al. (2018) reviewed 80 studies related to dynamic DEA
published in 1996 to 2016 through the elements of literature phases, structuring
models and applications.
For the ratio form model, in addition to the two cases introduced in Sect. 15.1,
Khalili-Damghani et al. (2015) developed a model to measure the efficiency of a
DMU through the whole planning horizon, in that partially consumed inputs accu-
mulate through the different time periods at a transfer rate of (1-β). In other words,
(1-β)t of the amount for each unit of a quasi-fixed input are retained in period t. The
proposed model was applied to 11 cotton producers in Iran, with fertilizer, seed, and
pesticide as the inputs, and cotton as the output, and with fertilizer and seed as
carryovers, being transferred from one period to the next. Zhang (2019) used a
two-stage model to evaluate the performance of 30 provincial high-tech industries in
China from 2013 to 2015 allowing the technology of different periods to change.
Alizadeh et al. (2020) built a network DEA model to assess the performance of
16 regional electricity companies. A linear combination of the efficiencies in suc-
cessive periods was used as the complement of the system. An et al. (2020) proposed
a two-stage approach that is able to take the lag effect into account to measure the
efficiency of the high-tech industries in 29 Chinese regions from 2010 to 2012. Yu
et al. (2021b) examined the innovation performance of 506 high-tech companies in
China, and the innovation process was decomposed into an R&D stage and a
commercialization stage.
For the distance function model, Silva and Stefanou (2007) proposed a hyperbolic
model to measure the long-term efficiency of a system, where the carryover (gross
investment) was treated as an output. They developed models in the context of an
adjustment-cost technology and inter-temporal cost minimization, and applied them
to measure the performance of 61 Pennsylvania dairy operations in the US for the
period 1986–1992. Jaenicke (2000) studied the effects of crop rotation, which
generates higher yields than growing identical crops in isolated monocultures. A
dynamic model was proposed to measure the output efficiency of a rotation by using
16 observations from an experimental farm in Pennsylvania, where three crops,
small grain, corn, and soybean, made up a rotation, and the organic matter in the soil
played the role of carryover. Chen (2009) extended the dynamic structure of the
whole-unit system in a period discussed in Färe and Grosskopf (1996) to network
systems. The author defined the dynamic technical efficiency for each division as the
ratio of the minimum input that must be used over the planning horizon to the total
input that has been used by the division, and the dynamic technical efficiency of a
DMU as the product of those of its component divisions. Aparicio and Kapelko
(2019) extended the measurement of dynamic inefficiency in the full input-output
space in the adjustment-cost framework to account for slacks and developed a

dynamic weighted additive model in the directional distance function form to
evaluate the efficiency of 311 large dairy firms in six EU countries from 2005 to
2012. Lin and Liu (2021) extended the multiplier dynamic DEA by using directional
distance function to measure the efficiencies of 48 mutual funds in the American
market from January 2016 to December 2018.
Bogetoft et al. (2009) studied the impact of public capital on technology and
productivity, using a division (or period, in the context of a dynamic system)
distance measure model. The state-level data of the US manufacturing sector for
the years 1978 to 1999 was used for illustration, where the carryovers were private
and public investments. Emrouznejad and Thanassoulis (2005) applied an input
division distance measure (one parameter for each period) to check whether an
assessment path is dynamically efficient. The distance parameters were attached to
both the inputs and carryovers, and the carryovers were represented by stock inputs.
The model was applied to 15 universities in the UK from 1994 to 1998.
Jahanshahloo et al. (2006) stated that every division distance parameter must be
less than or equal to one to be able to obtain efficient paths. Skevas et al. (2014)
modified the model of Skevas et al. (2012) by employing undesirable outputs (i.e.,
fertilizer, herbicides, and fungicides) as risk-adjusted inputs. The production tech-
nology assumes weak disposability of undesirable outputs. Fukuyama and Weber
(2015) constructed a three-year directional distance function to measure the effi-
ciency of 96 regional and five city banks in Japan. The operations in each period
were separated into two stages, deposit collection and investment, with assets, firm
loans, and securities as carryovers.
Lu et al. (2014) applied an input-oriented SBM model to evaluate the perfor-
mance of 34 life insurance companies in China for the period 2006–2010, where
debt and equity capital were treated as carryovers. Hung et al. (2014) proposed a
modified output-oriented SBM model to measure the efficiency of 109 semiconduc-
tor firms in Taiwan over the period 2006–2012, where employees were employed to
generate revenues and market value in each period, with liabilities and the owners’
equity as carryovers. The overall efficiency over the horizon was derived as a
weighted harmonic average of the period-specific efficiency. Alperovych et al.
(2013) analyzed the efficiency of 88 private equity backed leverage buyouts com-
pleted and exited during the period 1999–2008 in the UK, with shareholders’ equity
as the carryover. Avkiran and Goto (2011) used an SBM model to maximize the
slacks adjusted by ranges of observations to evaluate the overall efficiencies and
annual inefficiencies of 52 Japanese regional banks for the period 1998–2007, where
dividends and retained earnings were carryovers. Moreno et al. (2013) proposed a
combined SBM model to measure the performance of 26 US wireline telecommu-
nication companies from 1997 to 2007. Bai et al. (2015) applied the combined SBM
model of Tone and Tsutsui (2014) to measure the performance of 54 hi-tech zones in
China over the period 2007–2012, in which the system in each period was composed
of the production and R&D divisions, with capital and R&D staff as the carryovers.
The same model was also used by Chao et al. (2015) to measure the efficiency of
27 Taiwanese banks during the seven-year period 2005–2011. The system in each
period was separated into three divisions, capability, efficiency, and profitability, and
the non-performing loans and loan loss reserves were defined as carryovers for the
efficiency and profitability divisions, respectively.
There are many other applications that use the slacks-based measures. Yeh et al.
(2020) developed a three-stage approach that incorporates the carry-over effects of
both human capital and organizational forgetting into a dynamic SBM model to
evaluate the sustainability performance of 16 major industrial sectors in Taiwan for
the period 2010–2015. Chang et al. (2021) measured the multi-period efficiency of a
portfolio comprising 44 textile companies in Taiwan from years 2011 to 2018 by
nesting an additive model within an SBM model to utilize the information in
individual periods. Yu et al. (2021a) dealt with the heterogeneity problem related
to the determination of meta-technology when measuring performance in a dynamic
setting, and proposed a two-stage model to measure the efficiency of 22 Taiwanese
banks from 2008 to 2016. Li et al. (2022) proposed a Malmquist DEA model with
worst practice frontier to assess the bankruptcy risk of 4426 US banks over the
period of 2002–2016. Lu et al. (2022) assessed the performance of efficiency and
total factor productivity of 47 publicly-traded companies in Taiwan’s IC design
industry. Huang et al. (2022) proposed a modified SBM model to measure the
overall, production, and service efficiencies of 310 forest parks in China at a
provincial level during the period of 2009 to 2018.
In a series of works, Sengupta (1994a, 1994b, 1995, 1999) introduced capital
inputs along with variable inputs and generalized the DEA framework by capturing
the intertemporal effects of capital on outputs. The output is a function of capital
inputs and technical innovation that may vary over time due to non-uniform capital
utilization. Nemoto and Goto (2003) applied the model from Nemoto and Goto
(1999) to measure the performance of nine privately owned Japanese electric utilities
(with 135 observations) during the period 1981–1995, and Sueyoshi and Sekitani
(2005) extended the model from constant returns to scale to variable returns to scale.
Soleimani-damaneh (2009, 2013a, 2013b) proposed a number of solution methods
for the model of Nemoto and Goto (1999) when the production possibility set is a
free disposal hull. Ouellette and Yan (2008) proposed a model similar to that of
Nemoto and Goto (1999), although the restrictions on capital investment were
weaker. De Mateo et al. (2006) used profit maximization as the objective function
and examined 35 Chilean retail department stores for the years 2000 and 2001. The
results identified an optimal path of adjustment for the input quantities, and provided
a measure of the potential cost savings. Färe (1986) developed a model of maximiz-
ing a single output to measure the output efficiency in a dynamic environment which
allows for input allocation over finitely many periods. Färe et al. (2011) separated the
single output into final output and intermediate investment. Given that the objective
is to maximize the sum of the final output, the authors were able to obtain the
allocation of the final output and investment. Amirteimoori (2006) constructed a
model to measure the revenue efficiency of 11 gas companies in Iran during 2003
and 2004. The optimal output and carryover were calculated using the prices of the
output and the carryover (represented by the revenue of gas sold). Based on this, the
revenue efficiency of each period was calculated, and the efficiency of the whole
period was a weighted average of those of individual ones, where the weight was the
optimal revenue of each period in that of all periods. Kapelko et al. (2014) proposed
a model to measure the shadow value of quasi-fixed factors, based on which the
dynamic cost inefficiency was measured. The overall dynamic cost inefficiency
reflects a firm’s ability to minimize the production cost at a given level of outputs.
References
Akbarian, M., Najafi, E., Tavakkoli-Moghaddam, R., & Hosseinzadeh-Lotfi, F. (2015). A network-
based data envelopment analysis model in a dynamic balanced score card. Mathematical
Problems in Engineering, 2015, 1–13. https://doi.org/10.1155/2015/914108
Alizadeh, R., Beiragh, R. G., Soltanisehat, L., Soltanzadeh, E., & Lund, P. D. (2020). Performance
evaluation of complex electricity generation systems: A dynamic network-based data envelop-
ment analysis approach. Energy Economics, 91, 104894.
Alperovych, Y., Amess, K., & Wright, M. (2013). Private equity firm experience and buyout vendor
source: What is their impact on efficiency? European Journal of Operational Research, 228,
601–611.
Amirteimoori, A. (2006). Data envelopment analysis in dynamic framework. Applied Mathematics
and Computation, 181, 21–28.
An, Q., Meng, F., Xiong, B., Wang, Z., & Chen, X. (2020). Assessing the relative efficiency of
Chinese high-tech industries: A dynamic network data envelopment analysis approach. Annals
of Operations Research, 290, 707–729.
Aparicio, J., & Kapelko, M. (2019). Accounting for slacks to measure dynamic inefficiency in data
envelopment analysis. European Journal of Operational Research, 278, 463–471.
Avkiran, N. K. (2015). An illustration of dynamic network DEA in commercial banking including
robustness test. Omega, 55, 141–150.
Bai, X. J., Yan, W. K., & Chiu, Y. H. (2015). Performance evaluation of China’s Hi-tech zones in
the post financial crisis era - analysis based on the dynamic network SBM model. China
Economic Review, 34, 122–134.
Bogetoft, P., Färe, R., Grosskopf, S., Hayes, K., & Taylor, L. (2009). Dynamic network DEA: An
illustration. Journal of the Operations Research Society of Japan, 52, 147–162.
Chang, T. S., Tone, K., & Wu, C. H. (2021). Nested dynamic network data envelopment analysis
models with infinitely many decision making units for portfolio evaluation. European Journal
Chao, C. M., Yu, M. M., & Wu, H. N. (2015). An application of the dynamic network DEA model:
The case of banks in Taiwan. Emerging Markets Finance & Trade, 51, S133–S151.
Chen, C. M. (2009). A network-DEA model with new efficiency measures to incorporate the
dynamic effect in production networks. European Journal of Operational Research, 194,
687–699.
Chen, P. C. (2012). Measurement of technical efficiency in farrow-to-finish swine production using
multi-activity network data envelopment analysis: Evidence from Taiwan. Journal of Produc-
tivity Analysis, 38, 319–331.
Chen, C. M., & van Dalen, J. (2010). Measuring dynamic efficiency: Theories and an integrated
methodology. European Journal of Operational Research, 203, 749–760.
De Mateo, F., Coelli, T., & O'Donnell, C. (2006). Optimal paths and costs of adjustment in dynamic
DEA models: With application to Chilean department stores. Annals of Operations Research,
145, 211–227.
References 417
Emrouznejad, A., & Thanassoulis, E. (2005). A mathematical model for dynamic efficiency using
data envelopment analysis. Applied Mathematics and Computation, 160, 363–378.
Fallah-Fini, S., Triantis, K., & Johnson, A. (2014). Reviewing the literature on non-parametric
dynamic efficiency measurement: State-of-the-art. Journal of Productivity Analysis, 41, 51–67.
Färe, R. (1986). A dynamic non-parametric measure of output efficiency. Operations Research
Letters, 5, 83–85.
Färe, R., & Grosskopf, S. (1996). Intertemporal production frontiers: with dynamic DEA. Kluwer.
Färe, R., Grosskopf, S., Fukuyama, H., & Margaritis, D. (2011). DEA and endogenous technolog-
ical change. European Journal of Operational Research, 210, 457–458.
Fukuyama, H., & Weber, W. L. (2015). Measuring Japanese bank performance: A network DEA
approach. Journal of Productivity Analysis, 44, 249–264.
Avkiran, N. K., & Goto, M. (2011). A tool for scrutinizing bank bailouts based on multi-period peer
benchmarking. Pacific-Basin Finance Journal, 19, 447–469.
Huang, X. J., An, R., Yu, M. M., & He, F. F. (2022). Tourism efficiency decomposition and
assessment of forest parks in China using dynamic network data envelopment analysis. Journal
of Cleaner Production, 363, 132405.
Hung, S. W., He, D. S., & Lu, W. M. (2014). Evaluating the dynamic performances of business
groups from the carry-over perspective: A case study of Taiwan’s semiconductor industry.
Omega, 46, 1–10.
Jaenicke, E. C. (2000). Testing for intermediate outputs in dynamic models: Accounting for soil
capital in rotational crop production and productivity measures. Journal of Productivity Anal-
ysis, 14, 247–266.
Jahanshahloo, G. R., Soleimani-damaneh, M., & Reshadi, M. (2006). On Pareto (dynamically)
efficient paths. International Journal of Computer Mathematics, 83, 8–9.
Kao, C. (2014). Efficiency decomposition in network data envelopment analysis with slacks-based
Kapelko, M., Lansink, A. O., & Stefanou, S. E. (2014). Assessing dynamic inefficiency of the
Spanish construction sector pre- and post-financial crisis. European Journal of Operational
Research, 237, 349–357.
Khalili-Damghani, K., Tavana, M., Santos-Arteaga, F. J., & Mohtasham, S. (2015). A dynamic
multi-stage data envelopment analysis model with application to energy consumption in the
cotton industry. Energy Economics, 51, 320–328.
Li, Z., Feng, C., & Tang, Y. (2022). Bank efficiency and failure prediction: A nonparametric and
dynamic model based on data envelopment analysis. Annals of Operations Research, 315,
279–315.
Lin, R., & Liu, Q. (2021). Multiplier dynamic data envelopment analysis based on directional
distance function: An application to mutual funds. European Journal of Operational Research,
293, 1043–1057.
Lu, C. C., Chen, X., Lee, S. W., Hsu, S. Y., & Chou, K. W. (2022). Two stage’s performance
analysis of Taiwanese IC design industry: A dynamic network slacks-based data envelopment
analysis approach. International Journal of Technology Management, 89, 93–123.
Lu, W. M., Wang, W. K., & Kweh, Q. L. (2014). Intellectual capital and performance in the Chinese
life insurance industry. Omega, 42, 65–74.
Mariz, F. B. A. R., Almeida, M. R., & Aloise, D. (2018). A review of dynamic data envelopment
analysis: State of the art and applications. International Transactions in Operational Research,
25, 469–505.
Moreno, P., Lozano, S., & Gutiérrez, E. (2013). Dynamic performance analysis of US wireline
telecommunication companies. Telecommunications Policy, 37, 469–482.
Nemoto, J., & Goto, M. (1999). Dynamic data envelopment analysis: Modeling intertemporal
behavior of a firm in the presence of productive inefficiencies. Economics Letters, 64, 51–56.
Nemoto, J., & Goto, M. (2003). Measurement of dynamic efficiency in production: An application
of data envelopment analysis to Japanese electric utilities. Journal of Productivity Analysis, 19,
191–210.
Ouellette, P., & Yan, L. (2008). Investment and dynamic DEA. Journal of Productivity Analysis,
29, 235–247.
Sengupta, J. K. (1994a). Measuring dynamic efficiency under risk aversion. European Journal of
Sengupta, J. K. (1994b). Evaluating dynamic efficiency by optimal control. International Journal of
Systems Science, 25, 1337–1353.
Sengupta, J. K. (1995). Dynamic of data envelopment analysis: Theory of systems efficiency.
Kluwer.
Sengupta, J. K. (1999). A dynamic efficiency model using data envelopment analysis. International
Journal of Production Economics, 62, 209–218.
Silva, E., & Stefanou, S. E. (2007). Dynamic efficiency measurement: Theory and application.
American Journal of Agricultural Economics, 89, 398–419.
Skevas, T., Lansink, A. O., & Stefanou, S. E. (2012). Measuring technical efficiency in the presence
of pesticide spillovers and production uncertainty: The case of Dutch arable farms. European
Skevas, T., Stefanou, S. E., & Lansink, A. O. (2014). Pesticide use, environmental spillovers and
efficiency: A DEA risk-adjusted efficiency approach applied to Dutch arable farming. European
Soleimani-damaneh, M. (2009). An effective computational attempt in DDEA. Applied Mathemat-
ical Modelling, 33, 3943–3948.
Soleimani-damaneh, M. (2013a). An enumerative algorithm for solving nonconvex dynamic DEA
models. Optimization Letters, 7, 101–115.
Soleimani-damaneh, M. (2013b). Another approach for estimating RTS in dynamic DEA. Journal
of Productivity Analysis, 39, 75–81.
Sueyoshi, T., & Sekitani, K. (2005). Returns to scale in dynamic DEA. European Journal of
3–4.
Tone, K., & Tsutsui, M. (2014). Dynamic DEA with network structure: A slacks-based measure
Von Geymueller, P. (2009). Static versus dynamic DEA in electricity regulation: The case of US
transmission system operators. Central European Journal of Operations Research, 17,
397–413.
Yeh, L. T., Tseng, M. L., & Lim, M. K. (2020). Assessing the carry-over effects of both human
capital and organizational forgetting on sustainability performance using dynamic data envel-
opment analysis. Journal of Cleaner Production, 250, 119584.
Yu, M. M., Lin, C. I., Chen, K. C., & Chen, L. H. (2021a). Measuring Taiwanese bank perfor-
mance: A two-system dynamic network data envelopment analysis approach. Omega, 98,
102145.
Yu, A., Shi, Y., You, J., & Zhu, J. (2021b). Innovation performance evaluation for high-tech
companies using a dynamic network data envelopment analysis approach. European Journal of
Zhang, L. Y. (2019). Dynamic network data envelopment analysis based upon technology changes.
Infor, 57, 242–259.
Chapter 16
Linkage Efficiency
The feature that fundamentally characterizes a network system is the intermediate

products produced and consumed in the system. In the corresponding whole-unit
(black-box) system, only the (exogenous) inputs sent into the system to be used for
production and the (exogenous) outputs which are produced and sent out of the
system are taken into consideration when measuring the efficiency of the system. In
the case of a network system, on the other hand, the operations of the system’s
component divisions are factored into the efficiency assessment, so that the inter-
mediate products that are their inputs and outputs are taken into account. In short, a
network-based approach to evaluating the efficiency of a system incorporates both
exogenous and endogenous inputs and outputs. When the divisions of a system are
operating independently, a division that produces an intermediate product wishes to
produce more of it in order to maximize efficiency. On the contrary, the division that
consumes this intermediate product in its own production process wishes to consume
less of it in order to be more efficient. At times, there is a discrepancy between the
amount of the intermediate product produced by the supply division and that
consumed by the demand division. If the amount produced is greater than the
amount consumed, then the excess amount is left unused in the system and wasted.
In the preceding chapters, it was shown that the efficiency of the basic series
system, in radial measures, can be expressed as the product of the efficiencies of the
component divisions. The efficiency of the parallel system can be expressed as a
weighted average of the efficiencies of the component divisions. For general net-
work systems, those without feedback can be converted into a series of subsystems
composed of divisions that are connected in parallel. Based on the converted series-
parallel structures the relationship between the efficiencies of the system and of its
divisions are explored. The divisions that have the greatest impact on the perfor-
mance of the system as a whole can thus be identified. Making improvements to
these divisions will increase the efficiency of the entire system in the most effective
manner. Note that in these decompositions of efficiency, an assumption that the
intermediate products produced by the supply division are completely consumed by
https://doi.org/10.1007/978-3-031-27593-7_16
420 16 Linkage Efficiency
the demand division was implicitly made. It ignores the possibility that a portion of
the intermediate product is left unused. The situation in which the entire amount is
completely used up is a cooperative case, whereas the situation in which only a
portion of the amount is consumed is a non-cooperative case. When the intermediate
product linking the supply and demand divisions is not completely consumed, the
extent to which the intermediate product is being utilized is termed linkage efficiency
because the unused portion is wasted. This chapter discusses the measurement of
linkage efficiencies.
Although a radial measure of efficiency has the merit of being decomposable, it
also has the weakness of not being able to assign appropriate efficiency scores to
weakly efficient DMUs and those using weakly efficient DMUs as targets to
calculate efficiencies. Moreover, the decomposition works only under constant
returns to scale in most cases. In contrast, non-radial efficiency measures, of which
the slacks-based measure (SBM) of efficiency is the typical example, do not have the
problem related to the non-Archimedean number, and are more representative of the
performance of the DMU being assessed. The discussion presented in this chapter
thus elucidates the decomposition of the SBM efficiency. The discussion is divided
into three sections: Sect. 16.1 for basic series systems, Sect. 16.2 for parallel
systems, and Sect. 16.3 for general network systems.
16.1 Basic Series Systems
The basic series system is composed of a number of divisions connected in a

sequence, where the first division consumes all the exogenous inputs supplied
from the outside to produce only the intermediate products that are used by the
second division for the production of its own intermediate products, which in turn
are sent on to be consumed by the next division, and so on. This process is continued
until the last division in the system which uses all the intermediate products
produced by the division immediately preceding it to produce the final exogenous
outputs to be sent out of the system.
We start the discussion with the simplest case of the two-division system. We
then follow this by generalizing the concept to the multi-division system.
16.1.1 Two-division Systems
Consider the simplest case of a closed series system consisting of two divisions (see
Fig. 16.1). The inputs Xi, i = 1,. . ., m are supplied from the outside to Division
1 which produces the intermediate products Zg, g = 1,. . ., h, all of which are used by
Division 2 to produce the final outputs Yr, r = 1,. . ., s which are sent out of the
system. Suppose there are n DMUs whose efficiencies are to be measured. The
relational model for such a system under the assumption of constant returns to scale
16.1 Basic Series Systems 421

closed series system with
two divisions Xi Zg Yr
1 2
i=1,…, m g=1,…, h r=1,…, s
(see Sect. 10.2.1 in Chap. 10) shows that the system efficiency ES and the two
division efficiencies E(1) and E(2) of the DMU being assessed are ES = sr = 1 ur Y r0 /

m
i = 1 vi X i0 , E
(1)
= hg = 1 wg Z g0 / m
i = 1 vi X i0 , and E
(2)
= sr = 1 ur Y r0 / hg = 1 wg Z g0 .
The system efficiency can thus be decomposed into the product of the two division
efficiencies: ES = E(1) × E(2). This decomposition is based on the implicit assumption
that all the intermediate products produced by Division 1 are used up by Division
2. This model does not allow Division 2 to use less than the total amount produced
by Division 1, nor does it permit Division 1 to produce more than what is needed by
Division 2. In short, Divisions 1 and 2 are required to cooperate with each other.
In an effort to relax this restriction, Kao (2019) proposed using the SBM Model
(10.21) to measure efficiency:
1 m
1- s - =X i0
i=1 i
ρ = min:
S m
1 s
1þ sþ =Y r0
r=1 r
s
n
ð1Þ
s:t: λj X ij þ si- = X i0 , i = 1, . . . , m
j=1
n n
ð1Þ ð2Þ
λj Z gj ≥ λj Z gj , g = 1, . . . , h
j=1 j=1
n
ð2Þ
λj Y rj - sþ
r = Y r0 , r = 1, . . . , s
j=1
ð1Þ ð2Þ
λj , λj , si- , sþ
r ≥ 0, j = 1, . . . , n, i = 1, . . . , m, r = 1, . . . , s:
ð16:1Þ
ð1Þ ð2Þ
At optimality, nj = 1 λj Z gj and nj = 1 λj Z gj are the target outputs for Division
1 and target inputs for Division 2, respectively. Let
n n
ð1Þ ð2Þ
tþ
g = λj Z gj - Z g0 , t g- = Z g0 - λj Z gj , g = 1, . . . , h ð16:2Þ
j=1 j=1
denote the output shortage that should have been produced by Division 1 and the
excess input that can be saved by Division 2, respectively, if the two divisions are
operating efficiently.
Based on the inputs and outputs of the two divisions, their efficiencies can be
defined as:
h -
1- 1 m -
i = 1 si =X i0
1- 1
g = 1 t g =Z g0
ρð1Þ = m
, ρð 2 Þ = h
s þ
: ð16:3Þ
1 þh
1 h
g = 1 t g =Z g0
þ 1 þ 1s r = 1 sr =Y r0
Note that, in Eq. (16.2), it is possible that the current amount of the intermediate
product Zg0 exceeds the target output expected to be produced by Division 1, with t þg
being a negative number. Similarly, it is also possible that Zg0 is smaller than the
target input to be used by Division 2, with t g- being a negative number. In this case,
the factor efficiencies 1/(1 + t þ -
g / Zg0) and 1 - t g / Zg0 can be greater than one,
implying that the corresponding division is super-efficient.
The product of the two division efficiencies can be expressed as follows:
1 m 1 h
1- s - =X i0 1- t - =Z g0
g=1 g
ð1Þ ð2Þ m i=1 i h
ρ ×ρ = ×
1 h 1 s
1þ t þ =Z g0
g=1 g
1þ sþ =Y r0
r=1 r
h s ð16:4Þ
1 m 1 h
1- -
s =X i0 1 - t - =Z g0
g=1 g
m i=1 i h
= × = ρS × π:
1 s þ 1 h
1þ s =Y r0
r=1 r
1þ t þ =Z g0
g=1 g
s h
The two fractionals after the second equal sign are obtained by interchanging the
denominators of the two division efficiencies. The first fractional after the second
equal sign is the system efficiency, denoted as ρS, and the second fractional is the
efficiency of an imaginary process that uses Zg to produce Zg, denoted as π. If the
amount that leaves the imaginary process equals what enters it (i.e., t þ -
g = t g =0),
h -
then this process is efficient, with an efficiency of π = (1 - h g = 1 t g =Z g0 )/
1
(1 + 1h hg = 1 t þ
g =Z g0 ) = 1. If, on the other hand, the amount entering the process is
smaller than the amount leaving it, then a loss of intermediate product has occurred
ð1Þ ð2Þ
according to the condition of nj = 1 λj Z gj > nj = 1 λj Z gj , which means that a
portion of what is produced by Division 1 is not used by Division 2. In this case, the
ð1Þ
sum of t þ - þ -
g and t g (see Eq. (16.2)) is positive on the basis of t g + t g =
n
j = 1 λj Z gj
ð2Þ þ
- n
j = 1 λj Z gj > 0 (or t g >- t g- ), which suggests that π = (1 - 1h h -
g = 1 t g =Z g0 )/
(1 + 1h hg = 1 t þg =Z g0) < (1 - h
1 h -
- 1h hg = 1 t g- =Z g0) = 1, an indication
g = 1 t g =Z g0)/(1
of inefficiency. This process of using Zg to produce Zg links the output of Division
1 to the input of Division 2. The corresponding efficiency π can thus be called
linkage efficiency. Since both ρS and π are bounded by the value of one, then it is true
that ρ(1) × ρ(2) ≤ 1 (see Expression (16.4)), which implies that ρ(1) and ρ(2) cannot be
super-efficient at the same time.
Model (16.1) is a non-cooperative model, which requires only that the interme-
diate products produced by Division 1 be equal to or greater than those needed by
ð1Þ ð2Þ
Division 2: nj = 1 λj Z gj ≥ nj = 1 λj Z gj . If a cooperative model is used, then both
sides of this equation are equal, suggesting that Division 1 and Division 2 must agree
on the amount to be produced by the former and consumed by the latter. According
to the definitions of t þ - þ -
g and t g formulated in Eq. (16.2), it can be derived that t g + t g
þ -
=0 or t g = - t g . Denote η , η , and η as the efficiencies of the system, Division
S (1) (2)
1, and Division 2, respectively, in the cooperative case. The product of the two
division efficiencies can be expressed as:
1 m 1 h
1- s - =X i0 1- t - =Z g0
g=1 g
ð1Þ ð2Þ m i=1 i h
η ×η = ×
1 h 1 s
1þ t þ =Z g0
g=1 g
1þ sþ =Y r0
r=1 r
h s
1 m 1 h
1- s - =X i0 1 - t - =Z g0
g=1 g
m i=1 i h
= ×
1 s þ 1 h
1þ s =Y r0 1 þ
r=1 r
t þ =Z g0
g=1 g
s h
1 m 1 h 1 m
1- -
s =X i0 1 - t - =Z g0
g=1 g
1- s - =X i0
m i=1 i h m i=1 i
= × = = ηS
1 s þ 1 h 1 s
1þ s =Y r0 1 -
r=1 r
t - =Z g0
g=1 g
1þ sþ =Y r0
r=1 r
s h s
ð16:5Þ
which is just the system efficiency. Here, the linkage efficiency π is equal to one.
In contrast to the non-cooperative model associated with Eq. (16.4), a linkage
efficiency equal to one is obtained when the two divisions in the system cooperate
with each other. This reflects the fact that the efficiency of the system is the product
of the efficiencies of its component divisions, which is consistent with the results
produced with radial efficiency measures. If the two divisions do not cooperate, the
intermediate product may not be fully used up. The division efficiencies ρ(1) and ρ(2)
are calculated based on the same amount of the intermediate product Zg0, which
obviously understates the optimal efficiencies that the divisions can achieve. The
system efficiency ρS, however, is calculated based on the optimal amount of the
intermediate products for the two divisions, which necessitates that this score be
adjusted by incorporating the linkage efficiency for it to equal the product of the two
Although the radial and SBM efficiencies in the cooperative case produce the
same type of decomposition, i.e., the system efficiency is the product of the division
efficiencies, the former works only under the technology of constant returns to scale.
The SBM model, on the other hand, works in the context of both constant and
variable returns to scale. The reason is that the technology of variable returns to scale
is simply described by that of constant returns to scale with the addition of convexity
n ðk Þ
constraints: j = 1 λj =1 and k = 1, 2. Adding the two convexity constraints
obviously has no effect on the decomposition presented in Eqs. (16.4) and (16.5).
( p–1)
Xi Z (1) Z (k–1) Z (k) Zg Yr
1
g
… g
k
g
… p
i=1,..., m g=1,…, h (1) g=h(k–2) +1,...,h(k–1) g =h(k–1) +1,...,h(k) g=h(p–2) +1,...,h(p–1) r=1,..., s
Fig. 16.2 Structure of the basic system with multiple divisions
16.1.2 Multi-division Systems
The efficiency decomposition discussed for the two-division system can be gener-
alized to multi-division systems similar to the basic series system discussed in Sect.
12.1 in Chap. 12. To maintain continuity, the structure of the system presented in
Fig. 16.2 is an exact copy of Fig. 12.1. In this system, Division 1 consumes the
inputs Xi, i = 1,. . ., m, which are supplied from outside the system, in order to
produce the intermediate products Z ðg1Þ , g = 1,. . ., h(1). Division 2 then consumes all
these h(1) intermediate products to produce the intermediate products Z ðg2Þ ,
g = h(1) + 1,. . ., h(2), which are then sent on to Division 3 to be used in its own
production process, and so on. This input-production-output process is repeated until
the final division, Division p, that uses the intermediate products Z gðp - 1Þ , g = h( p -
2)
+ 1,. . ., h( p - 1) produced by Division p - 1 in order to produce the final outputs
Yr, r = 1,. . ., s to be sent out of the system. The superscripts associated with the
intermediate products are actually not needed because we assume that each inter-
mediate product is produced by only one division and also consumed by only one
division. However, they are shown for easy identification of the divisions they are
associated with.
The relational model for measuring system efficiency is:
s
ur Y r0
ES = max: r =m1
vi X i0
i=1
hð1Þ m
ð1Þ
s:t: wg Z gj - vi X ij ≤ 0, j = 1, . . . , n
g=1 i=1
hðkÞ hðk - 1Þ
ðkÞ ðk - 1Þ
wg Z gj - wg Z gj ≤ 0, k = 2, . . . , p - 1, j = 1, . . . , n
ðk - 1Þ ðk - 2Þ
g=h þ1 g=h þ1
s hðp - 1Þ
ðp - 1Þ
ur Y rj - wg Z gj ≤ 0, j = 1, . . . , n
r =1 g=hðp - 2Þ
þ1
ur , vi , wg ≥ ε, 8r, i, g
At optimality, the system and division efficiencies can be calculated as:

s
r = 1 ur Y r0
ES = m
i = 1 vi X i0
hð1Þ ð1Þ
ð1Þ g = 1 wg Z g0
E = m
i = 1 vi X i0
hðkÞ ðkÞ
ðkÞ g = hðk - 1Þ þ1 wg Z g0
E = , k = 2, . . . , p - 1
hðk - 1Þ Z ðk - 1Þ
g=h ðk - 2Þ w
þ1 g g0
s
r = 1 ur Y r0
E ðpÞ = ðp - 1Þ
h ðp - 1Þ
g = hðp - 2Þ þ1 wg Z g0
The system efficiency can be decomposed into the product of the division
efficiencies: ES = pk = 1 EðkÞ . As is the case of the two-division system, the efficien-
cies calculated by means of this model are affected by the non-Archimedean number
ε. Furthermore, the decomposition works only under constant returns to scale.
Based on the model for the two-division system, the SBM model for the multi-
division system can be formulated as:
1 m
1- s - =X i0
i=1 i
ρ = min:
S m
1 s
1þ sþ =Y r0
r =1 r
s
n
ð1Þ
s:t: λj X ij þ si- =X i0 , i=1, ...,m
j=1
n n
λj Z gj ≥ λj Z gj , g=hðk -1Þ þ 1, ...,hðkÞ , k =1, ...,p-1
j=1 j=1 ð16:6Þ
n
ðpÞ
λj Y rj -sþ
r =Y r0 , r =1, ...,s
j=1
n
ðkÞ
λj =1, k =1, ...,p
j=1
ðkÞ
λj , si- , sþ
r ≥0, j=1, ...,n, k =1, ...,p, i=1, ...,m, r =1, ...,s,
n ðk Þ ðkÞ
where h(0) = 0. In this model, j = 1 λj Z gj is the target output of Division k for
ð Þ ðk Þ
intermediate product Z ðgkÞ , and n kþ1
j = 1 λj Z gj is the target input of Division k + 1 for
Z ðgkÞ . If a constant returns to scale technology is assumed, then the convexity
n ðk Þ
constraints j = 1 λj = 1 can simply be removed.
Let t ðgkÞþ and t ðgkÞ - denote the output shortage of Z ðgkÞ that should have been
produced by Division k and the excess input of Z ðgkÞ that can be saved by Division
k + 1, respectively, if Divisions k and k + 1 are operating efficiently. They can be
expressed as:
n n
ðk Þ ðk Þ ðk Þ ðkþ1Þ ðkÞ ðk Þ
λj Z gj - t ðgkÞþ = Z g0 , λj Z gj þ t ðgkÞ - = Z g0 : ð16:7Þ
j=1 j=1
As is the case for the two-division system, a negative value of t ðgkÞþ indicates that
ðk Þ
the current amount of the intermediate product Z g0 exceeds what is expected to be
produced by Division k, and a negative value of t ðgkÞ - is a sign that the current
ðk Þ
amount of Z g0 is smaller than what is expected to be used by Division k + 1.
ðk Þ ðk Þ
Therefore, the output efficiency of Z ðgkÞ , calculated as Z g0 /(Z g0 + t ðgkÞþ), and its input
ðk Þ ðk Þ
efficiency, calculated as - (Z g0 t ðgkÞ - )/Z g0 ,
may have values greater than one, a
phenomenon of super-efficiency.
By defining the efficiency of a division in a conventional way as the product of
the arithmetic average of the input factor efficiencies and the harmonic average of the
output factor efficiencies, the efficiencies of the p divisions are:
1 m
1- s - =X i0
i=1 i
ρ ð1Þ
= m
1 hð1Þ ð1Þ
1 þ ð1Þ t ð1Þþ =Z f 0
g=1 g
h
1 hðk - 1Þ ðk - 1Þ
1- ðk - 1Þ
t ðk - 1Þ - =Z g0
g = hðk - 2Þ þ1 g
ðkÞ
ρ = h - hðk - 2Þ , k = 2, . . . , p - 1
1 hðkÞ ðkÞ
1 þ ðkÞ t ðkÞþ =Z g0
h - hðk - 1Þ
1 hðp - 1Þ ðp - 1Þ
1 - ðp - 1Þ t ðp - 1Þ - =Z g0
g = hðp - 2Þ þ1 g
ρðpÞ = h - hðp - 2Þ :
1 s
1þ sþ =Y r0
r=1 r
s
ð16:8Þ
It is possible that the division efficiency is greater than one if some of the
associated factors are highly super-efficient.
The product of all p division efficiencies defined in Expression (16.8) is:
1 m
p 1- s - =X i0
i=1 i
ðk Þ m
ρ =
1 hð1Þ ð1Þ
k=1 1 þ ð1Þ t ð1Þþ =Z g0
g=1 g
h
1 h ð k - 1Þ ðk - 1Þ
p-1 1- ðk - 1Þ
t ðk - 1Þ - =Z g0
× h - hð k - 2 Þ
1 h ðk Þ ðk Þ
k=2 1 þ ðkÞ t ðkÞþ =Z g0
h - hðk - 1Þ
1 hðp - 1Þ ðp - 1Þ
1- ðp - 1Þ
t ðp - 1Þ - =Z g0
× h - hð p - 2 Þ :
1 s
1þ sþ =Y r0
r=1 r
s
Shifting the denominator of each fractional to the same position as that of the next
one and moving the denominator of the last fractional to the same position as that of
the first one results in the following expression:
m -
p
1 - m1 i = 1 si =X i0
ρð k Þ = s þ
k=1
1 þ 1s r = 1 sr =Y r0
p-1 h ðk Þ ðk Þ - ðk Þ
1- 1
hðkÞ - hðk - 1Þ g = hðk - 1Þ þ1 t g =Z g0
× : ð16:9Þ
h ðk Þ ðk Þþ ðk Þ
k=1 1 þ hðkÞ -1hðk - 1Þ g = hðk - 1Þ þ1 t g =Z g0
The fractional in the parentheses is simply the efficiency of the system, ρS. Each
fractional in the brackets is the efficiency of an imaginary process that uses Z ðgkÞ to
produce Z ðgkÞ . If this efficiency is equal to one, then the amount that flows out of the
process is the same as the amount flowing in, indicating that this process is efficient.
If it is less than one, then a loss has occurred.
n ðk Þ ðk Þ n ðkþ1Þ ðk Þ
Since j = 1 λj Z gj ≥ j = 1 λj Z gj , as is required in Model (16.6), the
ðk Þþ ðk Þ -
definitions of t g and t g in Expression (16.7) imply that t ðgkÞþ + t ðgkÞ - ≥ 0
(or t ðgkÞþ ≥ - t ðgkÞ - ). Each fractional in the brackets are therefore less than or equal to
one. In other words, if the target output for the preceding division is the same as the
target input for the current division, then no loss occurs between the two consecutive
divisions, and this linkage is efficient. By definition,
hðkÞ ðk Þ - ðk Þ
ðk Þ
1- 1
hðkÞ - hðk - 1Þ g = hðk - 1Þ þ1 t g =Z g0
π =
h ðk Þ ðk Þþ ðk Þ
1 þ hðkÞ -1hðk - 1Þ g = hðk - 1Þ þ1 t g =Z g0
is the efficiency of the kth imaginary process, which is also the efficiency of the
linkage between Divisions k and k + 1. Expression (16.9) can thus be expressed as:
p p-1
ρ ð k Þ = ρS π ðk Þ : ð16:10Þ
k=1 k=1
In the case of a cooperative model, the division that supplies an intermediate

product within the system is required to agree with the demand division on the
amount to be produced and consumed so that no waste occurs. Accordingly, the
ðk Þ ðkÞ ðkþ1Þ ðk Þ
inequality constraints nj = 1 λj Z gj ≥ nj = 1 λj Z gj are replaced by the equality
ðk Þ ðk Þ ðkþ1Þ ðkÞ
constraints nj = 1 λj Z gj = nj = 1 λj Z gj . The sum of t ðgkÞþ and t ðgkÞ - defined in
Eq. (16.7) then becomes zero. In this context, two cases have to be examined: t ðgkÞþ =
t gðkÞ - = 0 and t gðkÞþ = -t ðgkÞ - ≠ 0. The former indicates that the intermediate product
Z ðgkÞ functions efficiently in its role as both the output of Division k and the input of
Division k + 1. The latter, on the other hand, indicates that if one role is inefficient,
with a positive value, then the other role will be super-efficient, with a negative
value.
Similar to the Non-cooperative Model (16.6), the system efficiency of the coop-
- þ
erative model also has the form of (1 - m1 m i = 1 si =X i0 )/(1 + s
1 s
r = 1 sr =Y r0 ). The
corresponding division efficiencies are the same as those defined in Expression
(16.8). Using η(k) to denote the efficiency of the cooperative case, the product of
the p division efficiencies can also be expressed as:
1 m
p 1- s - =X i0
i=1 i
ðk Þ m
η =
1 hð1Þ ð1Þ
k=1 1 þ ð1Þ t ð1Þþ =Z g0
g=1 g
h
1 h ð k - 1Þ ðk - 1Þ
p-1 1- ðk - 1Þ
t ðk - 1Þ - =Z g0
× h - hðk - 2Þ
1 hðkÞ ðk Þ
k=2 1 þ ðkÞ t ðkÞþ =Z g0
h - hðk - 1Þ
1 h ð p - 1Þ ðp - 1Þ
1- ð p - 1Þ
t ðp - 1Þ - =Z g0
× h - hðp - 2Þ :
1 s
1þ sþ =Y r0
r=1 r
s
ðk Þ
1 h ðk Þ
1-
1 m
s - =X i0 p-1 1- t ðkÞ - =Z g0
= m i=1 i
× h - hðk - 1Þ
ðk Þ
:
1 s 1 hðkÞ ðk Þ
1þ sþ =Y r0 k=1 1 þ ðk Þ t ðkÞþ =Z g0
s r=1 r
h - hð k - 1 Þ g = hðk - 1Þ þ1 g
Since t ðgkÞþ = t ðgkÞ - = 0 (or t ðgkÞþ = - t ðgkÞ - ) in the case of the cooperative model,
replacing t gðkÞþ in the denominator of the fractional in the brackets with - t ðgkÞ -
16.2 Parallel Systems 429
makes the denominator equal to the numerator, which results in a value equal to one.
Consequently, the imaginary process is efficient, which also implies that the linkage
efficiency between two consecutive divisions is equal to one. The product of the
p division efficiencies thus becomes:
p p-1
ηðkÞ = ηS πð k Þ = η S : ð16:11Þ
k=1 k=1
The system efficiency is thus equal to the product of the division efficiencies. In
the non-cooperative model, the division efficiencies are calculated based on the
current amount of the intermediate products instead of their optimal amount, their
values are less than or equal to the optimal values they can achieve. However,
because the system efficiency is calculated based on the optimal amount of the
intermediate products associated with every division, its value is overstated and so it
must be adjusted by incorporating the linkage efficiency for it to equal the product of
all division efficiencies.
16.2 Parallel Systems
The measurement of the efficiency of a parallel system is discussed in Chap. 13.

Figure 13.3 depicts the most general form of a parallel system, which we repeat here
as Fig. 16.3 in order to maintain continuity. This system consists of p divisions, with
ðk Þ
each Division k using the inputs X i , i = m(k - 1) + 1,. . ., m(k) to produce the outputs
Y ðrkÞ , r = s(k - 1) + 1,. . ., s(k), with m(0) = 0, m( p) = m, s(0) = 0, and s( p) = s.

parallel system
Xi(1) Yr(1)
1
i = m(0) + 1,...,m(1) r = s(0) + 1,..., s(1)
Xi(2) Yr(2)
2
i = m(1) + 1,...,m(2) r = s(1) + 1,...,s(2)
.
.
.
Xi(p) Yr(p)
p
i = m(p–1) + 1,...,m(p) r = s(p–1) + 1,...,s(p)
In a parallel system, its component divisions operate independently and there is

no intermediate product connecting any two divisions. If we suppose that Yr0 =
p ðk Þ p ðk Þ
k = 1 ur Y r0 and Xi0 = k = 1 vi X i0 , the radial efficiency of this system can be
calculated via the following relational model:
s
ur Y r0
r=1
ES = max: m
vi X i0
i=1
sðkÞ mðkÞ
ðkÞ ðkÞ
s:t: ur Y rj - vi X ij ≤ 0, k = 1, . . . , p, j = 1, . . . , n
r = sðk - 1Þ þ1 i = mðk - 1Þ þ1
ur , vi ≥ ε, 8r, i,
which reproduces Model (13.6) in Chap. 13.

When a set of optimal solutions (u*, v*) is obtained, the efficiencies of the system
and divisions are calculated as:
s
ur Y r0
r=1
E =S
m
vi X i0
i=1
sðkÞ ð16:12Þ
ðkÞ
ur Y r0
r = sðk - 1Þ þ1
E ðkÞ = , k = 1, . . . , p
mðkÞ
ðkÞ
vi X i0
i = mðk - 1Þ þ1
ðk Þ ðk Þ

By defining the weight associated with Division k as ω(k) = m
i = mðk - 1Þ þ1 vi X i0 /
m
i = 1 vi X i0 ,
the system efficiency can be expressed as an average of the division
efficiencies weighted by ω(k):
s sðkÞ
ðkÞ
ur Y r0 p ur Y r0
r=1 r = sðk - 1Þ þ1
ES = m = m
vi X i0 k=1 vi X i0
i=1 i=1
mðkÞ sðkÞ
ðkÞ ðkÞ
p vi X i0 ur Y r0 p
i = mðk - 1Þ þ1 r = sðk - 1Þ þ1
= m × = ωðkÞ E ðkÞ
mðkÞ
k=1 vi X i0 ðkÞ
vi X i0 k=1
i=1
i = mðk - 1Þ þ1
16.2 Parallel Systems 431
This decomposition works only under constant returns to scale.

Moreover, the efficiencies are affected by the value assigned to the non-Archime-
dean number ε.
The SBM model for measuring the efficiency of the system under variable returns
to scale is:
1 p mðkÞ ðkÞ - ðkÞ

1- k=1
s
i = mðk - 1Þ þ1 i
=X i0
ρ = min:
S m
1 p sðkÞ ðkÞ
1þ k=1
sðkÞþ =Y r0
r = sðk - 1Þ þ1 r
s
n
ðkÞ ðkÞ ðkÞ - ðkÞ
s:t: λj X ij þ si = X i0 , i = mðk - 1Þ þ 1, . . . , mðkÞ , k = 1, . . . , p
j=1
n
ðkÞ ðkÞ ðkÞþ ðkÞ
λj Y rj - sr = Y r0 , r = sðk - 1Þ þ 1, . . . , sðkÞ , k = 1, . . . , p
j=1
n
ðkÞ
λj = 1, k = 1, . . . , p
j=1
ðkÞ ðkÞ - ðkÞþ
λj , s i , sr ≥ 0, j = 1, . . . , n, k = 1, . . . , p, i = 1, . . . , m, r = 1, . . . , s:
ð16:13Þ
The efficiency of Division k is:
mðkÞ ðk Þ - ðk Þ
ðk Þ
1- 1
mðkÞ - mðk - 1Þ i = mðk - 1Þ þ1 si =X i0
ρ = sðkÞ ðk Þþ ðk Þ
, k = 1, . . . , p: ð16:14Þ
1 þ sðkÞ -1sðk - 1Þ r = sðk - 1Þ þ1 sr =Y r0
A comparison of the system efficiency ρS, defined in Model (16.13), with

the division efficiency ρ(k), defined in Expression (16.14), suggests that the
former can be expressed as a linear combination of the latter, where the weight
associated with Division k is the ratio of the denominator of the efficiency of
this division, defined in Expression (16.14), to the denominator of the efficiency
of the system, defined in Model (16.13). Since each division consumes a
different number of inputs, m(k) - m(k - 1), this weight must be adjusted by the
ratio (m(k) - m(k - 1))/m. The final weight is:
sðkÞ ðkÞþ ðkÞ

mðkÞ - mðk - 1Þ 1 þ sðkÞ -1sðk - 1Þ r = sðk - 1Þ þ1 sr =Y r0
νðkÞ = sðkÞ ðkÞþ ðkÞ
, k = 1, . . . , p:
m 1 þ 1s p
=Y r0
k=1 r = sðk - 1Þ þ1 sr
The linear combination of the efficiencies of all p divisions with this set of
weights is thus:
1 sðkÞ ðk Þ
p p ðk Þ ðk - 1Þ 1þ sðkÞþ =Y r0
ðk Þ ðk Þ m -m sðkÞ - sðk - 1Þ r = sðk - 1Þ þ1 r
ν ρ = ×
m 1 p s ðk Þ ðk Þ
k=1 k=1 1þ k=1
sðkÞþ =Y r0
s
1 m ðk Þ ðk Þ - ðk Þ
1 - ðkÞ s =X i0
m - mðk - 1Þ i = mðk - 1Þ þ1 i
1 s ðk Þ ðk Þ
1 þ ðkÞ ðk - 1Þ sðkÞþ =Y r0
s -s r = sðk - 1Þ þ1 r
1 m ðk Þ ðk Þ - ðk Þ
p ðk Þ
-m ðk - 1Þ 1- s =X i0
=
m mðkÞ - mðk - 1Þ i = mðk - 1Þ þ1 i
ðk Þ
m 1 p s ðk Þ
k=1 1þ k=1
sðkÞþ =Y r0
s
1 p mðkÞ ðk Þ - ðk Þ
1- k=1
s
=X i0
= m = ρS :
1 p sðkÞ ðk Þ
1þ k=1
sðkÞþ =Y r0
s
ð16:15Þ
That is, the efficiency of the system is a linear combination of the efficiencies
of all its divisions. Note that the sum of the p weights ν(k) may not equal one. It
equals one only when all divisions have the same number of inputs and outputs, with
m(k) - m(k - 1) =m/p and s(k) - s(k - 1) = s/p for k = 1,. . ., p. Only under this
condition will the system efficiency be a weighted average of the division
efficiencies.
Compared to the radial efficiencies shown in Eq. (16.12), the SBM efficiencies
are not affected by the non-Archimedean number. Moreover, the decomposition
works under both constant and variable returns to scale. A final point to be noted is
that, because a parallel system does not have intermediate products, there is no
difference between the cooperative and non-cooperative models.
16.3 General Network Systems
In contrast to a basic series system, a general network system does not have a clearly
defined structure such that the concept of linkage efficiency can be discussed in a
systematic way. Fortunately, by introducing dummy divisions into the model, any
network system without feedback can be converted into a series of subsystems each
of which is composed of a number of divisions connected in parallel. Based on the
discussion of the efficiency decomposition of the basic series and parallel systems
presented in the preceding sections, the system efficiency of a general network
16.3 General Network Systems 433
system can be expressed as a function of the division efficiencies and the linkage
efficiencies.
Each Division k in a network system is associated with at most four types of
ðk Þ
factors: the exogenous input X i , i 2 I(k), the exogenous output Y ðrkÞ , r 2 O(k), the
ðk Þ
endogenous input Z f , f 2 M , and the endogenous output Z ðgkÞ , g 2 N(k). The
(k)
variables I(k), O(k), M(k), and N(k) are the index sets of the inputs used by Division k,
the outputs it produces, the intermediate products it uses, and the intermediate
products it produces, respectively. Figure 16.4 shows the structure of the general
network system. This figure is more precise than the network structure shown in
Fig. 9.2 because it removes the factors that do not exist instead of assigning a zero
ða,k Þ
value to them. Another difference between these two structures is that pa = 1 Z f
ðk Þ
and pb = 1 Z ðgk,bÞ in Fig. 9.2 are replaced by Z f and Z ðgkÞ in Fig. 16.4, respectively.
For the sake of simplifying the notation, we assume that each intermediate product is
produced by only one division and is used by only one other division. The ideas to be
discussed next can easily be extended to cases where one intermediate product is
produced and consumed by more than one division, with the notation being more
complicated.
If we suppose that the intermediate product Zg is produced by Division gs and
used by Division gd, then Z ðggs Þ and Z ðggd Þ are the same and the superscripts gs and gd
can thus be omitted. However, they are still used for easy identification of the
division being discussed. Suppose there are h intermediate products. The model
for measuring the efficiency of the system can be formulated as:

general network system Xi(k),i ∈ I (k)
Zf(k) Zg(k)
k
f ∈ M (k) g ∈ N (k)
Yr(k),r ∈ O (k)
1 p ðkÞ - ðkÞ
1- p k=1
s
i2I ðkÞ i
=X i0
jI ðkÞ j
k=1
ρS = min:
1 p ðkÞ
1þ p k=1
sðkÞþ =Y r0
r2OðkÞ r
jOðkÞ j
k=1
n
s:t: λj X i j þ si = X i0 , i 2 I ðkÞ , k = 1, . . . , p
j=1
n
ðkÞ ðkÞ ðkÞ
λj Y r j - srðkÞþ = Y r0 , r 2 OðkÞ , k = 1, . . . , p
j=1
n n
ðg Þ ðg Þ ðgd Þ ðgd Þ
λj s Z g js = λj Zg j , g = 1, . . . , h
j=1 j=1
n
ðkÞ
λj = 1, k = 1, . . . , p
j=1
λj , si , sr ≥ 0, j = 1, . . . ,n, k = 1, . . . ,p, i = 1, . . . ,m, r = 1, . . . , s:
ð16:16Þ
where | I(k) | and | O(k) | denote the numbers of elements in the corresponding sets. If
ðg Þ ðg Þ ðg Þ ðg Þ
the equality constraint nj = 1 λj s Z gj s = nj = 1 λj d Z gj d is replaced by the inequality
n ðg Þ ðg Þ
n ðg Þ ðg Þ
j = 1 λj Z gj ≥ j = 1 λj
s s d
constraint Z gj d , then it becomes a non-cooperative
model.
To measure the efficiency of Division k, we refer to the notation for the interme-
ðk Þ ðkÞ ðk Þ - ðk Þ
diate product shown in Fig. 16.4. If we suppose that nj = 1 λj Z fj + t f = Zf 0 , f 2
ðk Þ ðkÞ ðk Þ
M(k), and nj = 1 λj Z gj - t ðgkÞþ = Z g0 , g 2 N(k), the efficiencies of Division k and
Linkage Zg are:
1 ðkÞ- ðkÞ ðkÞ- ðkÞ

1- s =X i0 þ t =Z f 0
ðkÞ j I j þ j M ðkÞ j
ðkÞ i2I ðkÞ i f 2M ðkÞ f
ρ = , k =1, ...,p
1 ðkÞ ðkÞ
1 þ ðkÞ r2OðkÞ r
sðkÞþ =Y r0 þ t ðkÞþ =Z g0
g2N ðkÞ g
j O j þ j N ðkÞ j
1 ðg Þ
1- ðgÞ g2LðgÞ g
t ðgd Þ- =Z g0d
j L j
πðgs ,gd Þ = , g=1, ...,h:
1 ðgs Þþ ðgs Þ
1 þ ðgÞ ðgÞ t g =Z g0
jL j g2L
ð16:17Þ
where L(g) is the index set of the intermediate products Zg produced by Division gs
and consumed by Division gd.
The system and division efficiencies of any network system can be calculated
based on Model (16.16) and Expression (16.17). However, due to the wide range of
possible network structures, it is difficult to identify a single general expression that
suitably shows the relationship between the divisions. We thus use a specific
network system that is generic enough to illustrate the general process of efficiency
decomposition. The idea is to convert the system into an equivalent form consisting
of a number of parallel subsystems connected in a series. Based on the efficiency
decomposition of the series and parallel systems presented in the preceding sections,
the efficiency of the network system can be decomposed into the product of the
efficiencies of the subsystems that are structured in parallel structures divided by the
linkage efficiencies. The efficiencies of the subsystems are linear combinations of
those of the corresponding component divisions.
Consider a network system consisting of three divisions with the structure shown
ðk Þ
in Fig. 16.5 in which each Division k uses the inputs X i , i = m(k - 1) + 1,. . ., m(k) to
ðk Þ (k - 1)
produce the outputs Y r , r= s + 1,. . ., s , with m(0) = 0, m(3) = m, s(0) = 0, and
(k)
s = s. Furthermore, Divisions 1 and 2 also produce the intermediate products Z ðg1Þ ,

(3)
g = 1,. . ., h(1) and Z ðg2Þ , g= h(1) + 1,. . ., h, respectively, for Division 3 to consume in
its own production process. The SBM model for measuring the efficiency of this
system, based on Model (16.16), is:
Yr(1) , r = s(0) + 1,..., s(1) X i(3) , i = m( 2) + 1,..., m(3)
X i(1) Z g(1)
1
i = m( 0) + 1,..., m(1) g = 1,..., h (1)
Yr( 3)
3
Z g( 2 ) r = s ( 2) + 1,..., s ( 3)
X i( 2 )
2
i = m (1) + 1,..., m ( 2) g = h (1) +1,..., h
Yr( 2) , r = s (1) + 1,..., s ( 2)
Fig. 16.5 Structure of the example system

1 3 mðkÞ ðkÞ - ðkÞ

1- k=1
s
=X i0
ρ = min:
S m
1 3 sðkÞ ðkÞ
1þ k=1
sðkÞþ =Y r0
s
n
s:t: λj X ij þ si = X i0 , i = mðk - 1Þ þ 1, . . . , mðkÞ , k = 1, 2, 3
j=1
n
λj Y rj - sr = Y r0 , r = sðk - 1Þ þ 1, . . . , sðkÞ , k = 1, 2, 3
j=1
n n
λj Z gj = λj Z gj , g = 1, . . . , hð1Þ
j=1 j=1
n n
λj Z gj = λj Z gj , g = hð1Þ þ 1, . . . , h:
j=1 j=1
n
ðkÞ
λj = 1, k = 1, 2, 3
j=1
λj , si , sr ≥ 0, j = 1, . . . , n, k = 1, 2, 3, i = 1, . . . , m, r = 1, . . . , s:
ð16:18Þ
n ð1Þ ð1Þ
For a non-cooperative case, the equality constraints j = 1 λj Z gj =
n ð3Þ ð1Þ ð2Þ ð2Þ ð3Þ ð2Þ
j = 1 λj Z gj and nj = 1 λj Z gj = nj = 1 λj Z gj are replaced by the inequality
n ð 1Þ ð1 Þ n ðn3Þ ð1 Þ n ð2Þ ð2Þ ð3Þ ð2Þ
constraints j = 1 λj Z gj ≥ j = 1 λj Z gj
j = 1 λj Z gj
and≥ j = 1 λj Z gj ,
respectively.
Denoting the slack variables associated with the intermediate products as t ðg1Þ - ,
t gð1Þþ , t gð2Þ - , and t ðg2Þþ , we have:
n n
ð1Þ ð1Þ ð1Þþ ð1Þ ð3Þ ð1Þ ð1Þ - ð1Þ
λj Z gj - t g = Z g0 , λj Z gj þ t g = Z g0 , g = 1, . . . , hð1Þ
j=1 j=1
n n
ð2Þ ð2Þ ð2Þþ ð2Þ ð3Þ ð2Þ ð2Þ - ð2Þ
λj Z gj - t g = Z g0 , λj Z gj þ t g = Z g0 , g = hð1Þ þ 1, . . . , h
j=1 j=1
ðkÞþ ðkÞ -
tg , tg unrestricted in sign, k = 1, 2, g = 1, . . . , h:
The efficiencies of the three divisions and the two linkages, based on Expression
(16.17), are:
1 mð1Þ ð1Þ- ð1Þ

1- s =X i0
ρ ð1Þ
= mð1Þ -mð0Þ i=mð0Þ þ1 i
1 sð1Þ ð1Þ h ð 1Þ ð1Þ
1þ ð1Þ r=sð0Þ þ1
sðr1Þþ =Y r0 þ t ð1Þþ =Z g0
g=1 g
ðsð1Þ -sð0Þ Þþh
1 mð2Þ ð2Þ- ð2Þ
1- s =X i0
ρ ð2Þ
= mð2Þ -mð1Þ i=mð1Þ þ1 i
1 sð2Þ ð2Þ h ð2Þ
1þ sð2Þþ =Y r0 þ t ð2Þþ =Z g0
g=hð1Þ þ1 g
ðsð2Þ -sð1Þ Þþ h-hð1Þ r=sð1Þ þ1 r
mð3Þ ð3Þ- ð 3Þ hð1Þ ð1Þ- ð1Þ

1 s
i=mð2Þ þ1 i
=X i0 þ g=1 g
t =Z g0
1- ð3Þ
ðm -mð2Þ Þþh þ
h
g=hð1Þ þ1 g
t
ð2Þ- ð 2Þ
=Z g0
ρð3Þ =
1 s ð 3Þ ð3Þ
1þ sð3Þþ =Y r0
sð3Þ -sð2Þ r =sð2Þ þ1 r
1 hð1Þ ð1Þ- ð1Þ
1-
ð1Þ g=1
t g =Z g0
πð1,3Þ = h
1 hð1Þ ð1Þþ ð1Þ
1þ ð1Þ g=1 g
t =Z g0
h
1 h ð2Þ
1- ð1Þ
t ð2Þ- =Z g0
g=hð1Þ þ1 g
ð2,3Þ h-h
π = :
1 h ð2Þþ ð2Þ
1þ ð Þ t =Z
h-hð1Þ g=h þ1 g
1 g0
ð16:19Þ
To express the efficiency of the entire system as a function of the division

efficiencies and the linkage efficiencies, the original system with the structure
shown in Fig. 16.5 is transformed into one with the series-parallel structure shown
in Fig. 16.6, consisting of a series of two subsystems of three divisions each
I II
X i(1) Yr(1) Z g(1) , Z g( 2) Yr( 3)

1 3
Z (1)
g X i( 3)
X i(1) , X i( 2 ) , X i( 3) X i( 2 ) Y r( 2 ) Yr(1) Yr(1) Yr(1) , Yr( 2 ) , Yr( 3)
2 5
i = 1,..., m Z g( 2 ) r = 1,..., s
X i( 3) X i( 3) Yr( 2 ) Yr( 2 )
4 6
Fig. 16.6 Structure of the series-parallel transformation of the example system

connected in parallel. Divisions 4, 5, and 6, represented by circles, are dummy

divisions used for carrying information. They do not engage in actual production
ð3Þ
activities. More specifically, Division 4 transmits X i from outside the system
ð1Þ ð2Þ
into it, and Divisions 5 and 6 transmit Y r and Y r , respectively, from inside
the system to the outside. Note that each dummy division is limited to carrying
only one factor in order to make the transformed system equivalent to the
original system. The reason for this will become obvious in the discussion that
follows. The model for measuring the efficiency of this transformed system is
the same as Model (16.18) which was used for the original system, except that
the constraints corresponding to the dummy divisions must be taken into
consideration.
The constraints corresponding to Dummy Division 4 are:
n
μj X ij þ si = X i0 , i = mð2Þ þ 1, . . . , m
j=1
n
μj X ij - si = X i0 , i = mð2Þ þ 1, . . . , m:
j=1
ð3Þ
These two sets of constraints are formulated for Input X i from the input and
ð3Þþ ð3Þ -
output sides of this division. They are actually the same, with si =-si . The
first set of constraints is the same as the first set of constraints in Model (16.18) for
ð4Þ ð3Þ ð3Þ - ð3Þ - ð3Þþ ð3Þ -
k = 3, with μj = λj . We thus have si = si and si = - si . These two
sets of constraints are therefore redundant and not needed. The efficiency of this
dummy division can thus be expressed as:
mð3Þ ð3Þ - ð3Þ

ð4Þ
1- 1
mð3Þ - mð2Þ i = mð2Þ þ1 si =X i0
ρ = mð3Þ ð3Þþ ð3Þ
1 þ mð3Þ -1 mð2Þ i = mð2Þ þ1 si =X i0
mð3Þ ð3Þ - ð3Þ
1- 1
m ð 3Þ - m ð 2Þ i = mð2Þ þ1 si =X i0
= = 1,
m ð 3Þ ð3Þ - ð3Þ
1 þ mð3Þ -1 mð2Þ i = mð2Þ þ1 - si =X i0
which is efficient.
Similarly, the constraints corresponding to Dummy Divisions 5 and 6 are:
n
ð5Þ ð1Þ ð1Þ
μj Y rj þ sð1Þ
r
-
= Y r0 ,
j=1
n
ð5Þ ð1Þ ð1Þ
μj Y rj - sð1Þþ
r = Y r0 , r = sð0Þ þ 1, . . . , sð1Þ
j=1
n
ð6Þ ð2Þ ð2Þ
μj Y rj þ sð2Þ
r
-
= Y r0 ,
j=1
n
ð6Þ ð2Þ ð2Þ
μj Y rj - sð2Þþ
r = Y r0 , r = sð1Þ þ 1, . . . , sð2Þ ,
j=1
where the first two sets and last two sets of constraints correspond to the intermediate
products Y ðr1Þ and Y ðr2Þ in Divisions 5 and 6, respectively. The two constraints in each
set are the same and can be merged, resulting in sðr1Þ - =- srð1Þþ and sðr2Þ - =- sðr2Þþ .
Moreover, the two sets of constraints are the same as the second set of constraints in
Model (16.18) for k = 1 and k = 2, with μj = λj , sðr1Þþ = sðr1Þþ and μj = λj ,
ð2Þþ
sr = sðr2Þþ . All the constraints corresponding to these two dummy divisions are
thus redundant. If Y ðr1Þ and Y ðr2Þ are carried by one dummy division instead of two,
then they have the same intensity coefficients μj, while those for Y ðr1Þ and Y ðr2Þ in
ð1Þ
Model (16.18) are different. The intensity coefficient μj cannot equal both λj and
ð2Þ
λj at the same time, making it a non-redundant constraint which cannot be deleted.
The efficiencies of these two dummy divisions can be expressed as:
1 sð1Þ ð1Þ 1 sð1Þ ð1Þ

1- sð1Þ- =Y r0 1- -sð1Þþ =Y r0
ð5Þ
ρ = sð1Þ r=sð0Þ þ1 r
= s
ð1Þ r=sð0Þ þ1 r
=1
1 sð1Þ ð1Þþ ð1Þ 1 sð1Þ ð1Þ
1þ ð1Þ s =Y r0
r=sð0Þ þ1 r
1þ ð1Þ sð1Þþ =Y r0
r=sð0Þ þ1 r
s s
1- sð2Þ- =Y r0 1- -sð2Þþ =Y r0
ð6Þ
ρ = sð2Þ -sð1Þ r=sð1Þ þ1 r
= sð2Þ -sð1Þ r=sð1Þ þ1 r
=1
1þ sð2Þþ =Y r0 1þ sð2Þþ =Y r0
sð2Þ -sð1Þ r=sð1Þ þ1 r sð2Þ -sð1Þ r=sð1Þ þ1 r
which are also efficient.

Removing these redundant constraints makes the model corresponding to the
transformed system the same as of the model for the original system, indicating that
these two systems are equivalent.
The transformed system has two subsystems connected in a series, with
Subsystem I using X i , X i , and X i to produce Y ðr1Þ , Z ðg1Þ , Y ðr2Þ , Z ðg2Þ , and X i . Its
efficiency is thus expressed as:
1 3 mðkÞ ðkÞ - ðkÞ

1- k=1
s
i = mðk - 1Þ i
=X i0
ρ =I m
T
2 hðkÞ ðkÞþ 2 sðkÞ ðkÞþ
1 tg sr
T= 1þ þ
h þ sð2Þ þ ðmð3Þ - mð2Þ Þ k = 1 g = hðk - 1Þ þ1 Z g0
ðkÞ
k = 1 r = sðk - 1Þ þ1 Y r0
ðkÞ
mð3Þ ð3Þ -
ð - si Þ
þ ð3Þ
,
i = mð2Þ þ1 X i0
where h(0) = 0 and h(2) = h. This subsystem is composed of Divisions 1, 2, and

4 connected in parallel. Based on the conclusion obtained in Sect. 16.2 that the
system efficiency of the parallel system is a linear combination of the division
efficiencies, we have:
ρI = νð1Þ ρð1Þ þ νð2Þ ρð2Þ þ νð4Þ ρð4Þ , ð16:20Þ
where ρ(4) = 1, and
mð1Þ 1 s ð 1Þ ð1Þ hð1Þ ð1Þþ ð1Þ

1þ sð1Þþ =Y r0 þ t =Z f 0
m s þ hð1Þ
ð1 Þ r=1 r f =1 f
νð1Þ =
T
ð 2Þ
mð2Þ - mð1Þ 1 s
s
ð2Þþ ð 2Þ
=Y r0
1þ r = sð1Þ þ1 r
ðsð2Þ - sð1Þ Þ þ h - hð1Þ ð2Þþ ð2Þ

h
m þ t
g = hð1Þ þ1 g
=Z g0
νð2Þ =
T
ð3Þ ð2Þ
m -m 1 m ð3Þ - ð3Þ
1þ - si =X i0
ð4Þ
ν = m m - m ð2Þ i = mð2Þ þ1
:
T
ð3Þ
Subsystem II uses Z ðg1Þ , Z ðg2Þ , X i , Y ðr1Þ , and Y ðr2Þ to produce Y ðr3Þ , Y ðr1Þ , and Y ðr2Þ . Its
efficiency is thus expressed as:
2 hk t kg - 2 sk m3
- skþ s3i -
1 - hþs2 þm1 3 - m2 Z kg0
þ r
Y kr0
þ X 3i0
k = 1 g = hk - 1 þ1 k = 1 r = sk - 1 þ1 i = m2 þ1
ρII = :
3 sk
skþ
1þ 1
s
r
Y kr0
k = 1 r = sk - 1 þ1
This subsystem also has a parallel structure with three divisions, 3, 5, and
6. The subsystem efficiency is thus a linear combination of the three division
efficiencies:
ρII = νð3Þ ρð3Þ þ νð5Þ ρð5Þ þ νð6Þ ρð6Þ , ð16:21Þ
where ρ(5) = ρ(6) = 1, and
1 s ð3Þ
h þ mð3Þ - mð2Þ 1þ sð3Þþ =Y r0
ν ð3Þ
= s - sð2Þ r = sð2Þ þ1 r
h þ ðmð3Þ - mð2Þ Þ þ sð2Þ 1þ
1 3 sðkÞ ðk Þ
sðkÞþ =Y r0
s k=1 r = sðk - 1Þ þ1 r
1 sð1Þ ð1Þþ ð1Þ
1þ s =Y r0
sð1Þ sð1Þ r=1 r
νð5Þ =
1 3 sðkÞ ðk Þ
sðkÞþ =Y r0
s k=1 r = sðk - 1Þ þ1 r
1 s ð 2Þ ð2Þ
ð2Þ ð1Þ 1þ sð2Þþ =Y r0
ð6Þ s -s sð2Þ - sð1Þ r = sð1Þ þ1 r
ν = :
1 3 sðkÞ
sðkÞþ =Y r0
ðk Þ
s k=1 r = sðk - 1Þ þ1 r
Since the transformed system is composed of Subsystems I and II connected in a

series, we have ρI × ρII = ρSπ (I, II) based on the result obtained in Sect. 16.1 that the
product of the division efficiencies of a series system is the system efficiency
multiplied by the linkage efficiencies, where:
πðI,IIÞ
h ðk Þ sðkÞ m ð 3Þ
2 ðk Þ -
tg
2
ð- sðrkÞþ Þ si
ð3Þ -
1
hþsð2Þ þðmð3Þ - mð2Þ Þ ðk Þ þ ðk Þ þ ð 3Þ
k = 1 g = hðk - 1Þ þ1 Z g0 k = 1 r = sðk - 1Þ þ1 Y r0 X i0
i = mð2Þ þ1
= :
T
According to Fig. 16.6, the linkage between Subsystems I and II is an imaginary

ð3Þ
subsystem that uses X i , Y ðr1Þ , Y ðr2Þ , Z ðg1Þ , and Z gð2Þ to produce themselves, which
constitutes a parallel system composed of these factors. Accordingly, the linkage
efficiency π (I, II) is a linear combination of the efficiencies of these factors:
πðI,IIÞ = νðX 3 Þ ρðX 3 Þ þ νðY 1 Þ ρðY 1 Þ þ νðY 2 Þ ρðY 2 Þ þ νðZ 1 Þ ρðZ 1 Þ þ νðZ 2 Þ ρðZ 2 Þ , ð16:22Þ
where
1 m ð 3Þ ð3Þ - ð3Þ
1- s =X i0
ρ ðX 3 Þ
= mð3Þ - mð2Þ i = mð2Þ þ1 i
=1
1 mð3Þ ð3Þ - ð3Þ
1þ - s =X
mð3Þ - mð2Þ i = mð2Þ þ1 i i0
ð 1Þ
1 s ð1Þ
1 - ð1Þ r=1
- sðr1Þþ =Y r0
ρ ðY 1 Þ
= s =1
1 sð1Þ ð1Þþ ð1Þ
1 þ ð1Þ r=1
s r =Y r0
s
1 s ð 2Þ ð2Þ
1 - ð2Þ ð1Þ
- sðr2Þþ =Y r0
ρ ðY 2 Þ
= s - s r = sð1Þ þ1
=1
1 s ð 2Þ ð2Þþ ð2Þ
1 þ ð2Þ s =Y
s - sð1Þ r = s þ1 ð 1 Þ r r0
1 hð1Þ ð1Þ - ð1Þ

1 - ð1Þ g = 1
t g =Z g0
ρðZ 1 Þ = πð1,3Þ = h
1 hð1Þ ð1Þþ ð1Þ
1 þ ð1Þ g = 1
t g =Z g0
h
1 h ð2Þ
1- ð1Þ g = hð1Þ þ1 g
t ð2Þ - =Z g0
ðZ 2 Þ
ρ =π ð2,3Þ
= h - h
1 h ð2Þ
1þ ð1Þ = ð 1Þ
þ1
t ðg2Þþ =Z g0
h-h g h
mð3Þ - mð2Þ 1 m ð 3Þ ð3Þ - ð3Þ

ð3Þ - mð2Þ Þ þ sð2Þ
1 þ ð3Þ - mð2Þ
- si =X i0
h þ ð m m i = mð2Þ þ1
νðX 3 Þ =
T
ð1Þ sð1Þ ð1Þþ
s 1 ð1Þ
ð Þ ð Þ ð Þ
1 þ ð1Þ s =Y r0
ðY 1 Þ h þ ðm - m Þ þ s
3 2 2 s r =1 r
ν =
T
sð2Þ - sð1Þ 1 s ð 2Þ ð2Þ
ð3Þ - mð2Þ Þ þ sð2Þ
1 þ ð2Þ - sð1Þ
sð2Þþ =Y r0
h þ ð m s r = sð1Þ þ1 r
νðY 2 Þ =
T
ð1Þ hð1Þ ð1Þþ
h 1 ð1Þ
ð Þ ð Þ ð Þ
1 þ ð1Þ t =Z g0
h þ ð m 3 - m 2 Þ þ s 2
h g =1 g
νðZ 1 Þ =
T
h - hð1Þ 1 h ð2Þ
ð Þ ð Þ ð Þ
1þ ð Þ g = hð1Þ þ1 g
t ð2Þþ =Z g0
ðZ 2 Þ h þ ðm - m Þ þ s
3 2 2
h-h 1
ν = :
T
It should be noted that the efficiencies ρðX 3 Þ , ρðY 1 Þ , and ρðY 2 Þ are those of Divisions
4 (ρ(4)), 5 (ρ(5)), and 6 (ρ(6)), respectively, which are known to be efficient with a
score of one. The efficiencies ρðZ 1 Þ and ρðZ 2 Þ are the linkage efficiencies π (1, 3) and
π (2, 3), respectively. By incorporating Expressions (16.20), (16.21), and (16.22), ρS =
ρI × ρII / π (I, II) can be expressed as:
νð1Þ ρð1Þ þ νð2Þ ρð2Þ þ νð4Þ × νð3Þ ρð3Þ þ νð5Þ þ νð6Þ

ρS = :
νðX 3 Þ þ νðY 1 Þ þ νðY 2 Þ þ νðZ 1 Þ πð1,3Þ þ νðZ 2 Þ πð2,3Þ
In the case of the cooperative model where the intermediate products produced by
the supply division are the same as those used by the demand division, the linkage
efficiency π is equal to one. The decomposition thus becomes:
ρS = νð1Þ ρð1Þ þ νð2Þ ρð2Þ þ νð4Þ × νð3Þ ρð3Þ þ νð5Þ þ νð6Þ : ð16:23Þ
As a comparison, the radial efficiency of the system shown in Fig. 16.5 under
constant returns to scale can be calculated via the following relational model:
sð1Þ
ð1Þ sð2Þ ð2Þ sð3Þ ð3Þ
ur Y r0 þ r =sð1Þ þ1 ur Y r0 þ r =sð2Þ þ1 ur Y r0
r =sð0Þ þ1
ES = max:
mð1Þ
ð1Þ mð2Þ ð2Þ mð3Þ ð3Þ
vi X i0 þ i=mð1Þ þ1 vi X i0 þ i=mð2Þ þ1 vi X i0
i=mð0Þ þ1
ð1Þ ð1Þ ð1Þ
s:t: ur Y rj þ wg Z gj - vi X ij ≤0, j=1, ...,n
r =sð0Þ þ1 g=1 i=mð0Þ þ1
sð2Þ h mð2Þ
ð2Þ ð2Þ ð2Þ
ur Y rj þ wg Z gj - vi X ij ≤0, j=1, ...,n
r =sð1Þ þ1 g=h þ1 ð1Þ i=mð1Þ þ1
sð3Þ mð3Þ hð1Þ h

ur Y rj - vi X ij þ wg Z gj þ wg Z gj ≤0, j=1, ...,n
r =sð2Þ þ1 i=mð2Þ þ1 g=1 ð1Þ
g=h þ1
ur , vi , wg ≥ε, 8r,i,g
Based on the constraints, the system and division efficiencies at optimality are
calculated as:
s ð 1Þ ð1Þ s ð 2Þ ð2Þ sð3Þ ð3Þ

r = sð0Þ þ1 ur Y r0 þ r = sð1Þ þ1 ur Y r0 þ r = sð2Þ þ1 ur Y r0
E = S
m ð 1 Þ
ð1Þ þ m ð 2 Þ
ð2Þ þ m ð3 Þ
ð3Þ
i = mð0Þ þ1 vi X i0 i = mð1Þ þ1 vi X i0 i = mð2Þ þ1 vi X i0
s ð 1Þ ð1Þ h ð 1Þ ð1Þ
r = sð0Þ þ1 ur Y r0 þ g = 1 wg Z g0
E ð1Þ = m ð 1Þ ð1Þ
s ð 2Þ ð2Þ h ð2Þ
r = sð1Þ þ1 ur Y r0 þ g = hð1Þ þ1 wg Z g0
E ð2Þ = mð2Þ ð2Þ
sð3Þ ð3Þ
E ð3Þ =
m ð 3Þ ð3Þ h ð 1Þ ð1Þ h ð2Þ
i = mð2Þ þ1 vi X i0 þ g = 1 wg Z g0 þ g = hð1Þ þ1 wg Z g0
The same conversion that was carried out to transform the system in Fig. 16.5 to
the system in Fig. 16.6 can be applied to this relational model. The constraints
corresponding to the three dummy divisions are redundant, and the corresponding
efficiencies are equal to one. The efficiency of the entire system is then the product of
the efficiencies of Subsystems I and II, which are weighted averages of their
corresponding component divisions, with the weight associated with a given divi-
sion being the aggregate input consumed by this division in the aggregate input
consumed by all divisions in the same parallel subsystem.
Let
mð1Þ ð1Þ
ωð1Þ = m ð 1Þ ð1Þ þ mð2Þ ð2Þ þ mð3Þ ð3Þ
mð2Þ ð2Þ
mð3Þ ð3Þ
for the three divisions of Subsystem I, and
m ð 3Þ ð3Þ hð1Þ ð1Þ h ð2Þ

ð3Þ i = mð2Þ þ1 vi X i0 þ g = 1 wg Z g0 þ g = hð1Þ þ1 wg Z g0
ω =
S
s ð 1Þ ð1Þ
ωð5Þ =
S
s ð 2Þ ð2Þ
ωð6Þ =
S
mð3Þ hð1Þ h sð1Þ s ð 2Þ
S= vi X i0 þ wg Z g0 þ wg Z g0 þ ur Y r0 þ ur Y r0
i = mð2Þ þ1 g=1 ð 1Þ
g = h þ1 r = sð0Þ þ1 r = sð1Þ þ1
for the three divisions of Subsystem II. The system efficiency can then be
expressed as:
ES = ωð1Þ E ð1Þ þ ωð2Þ E ð2Þ þ ωð4Þ × ωð3Þ Eð3Þ þ ωð5Þ þ ωð6Þ :
This equation has exactly the same form as Eq. (16.23), which expresses the
decomposition of the SBM efficiency in the cooperative case, except that the weights
and efficiencies are calculated differently. The reason for this similarity is that in the
radial measure of efficiency, it is assumed that the intermediate products produced
by the supply division are used up by the demand division, leaving no waste in the
system, which is exactly the cooperative case.
References 445
The concept of linkage efficiency was introduced by Kao (2019) for basic series
systems, and it was later extended to general network systems by Kao (2020).
Galagedera et al. (2018) discussed a similar idea using radial efficiency measures.
They allowed the multipliers associated with the intermediate products produced by
the supply division to be different from those associated with the same intermediate
products consumed by the demand division, and they defined the ratio of the
aggregate value of the intermediate products of the demand division to that of the
supply division as a measure of linkage performance, which they called the internal
resource imbalance (IRI). To ensure that the IRI is positive, the multipliers
corresponding to the supply division must be greater than or equal to those
corresponding to the demand division. In another relevant study, Lozano and Khezri
(2021) applied the same idea as Kao (2019) to the measurement of the inefficiency of
the intermediate products in the case of the non-cooperative model. The inefficiency
of the system was decomposed into the sum of the inefficiencies of its component
divisions and those of the intermediate products. Finally, Zhao et al. (2022) used an
idea of coordination efficiency to measure the coordination effect that takes place in
a supply chain modelled as a two-stage system, which is similar to the idea of linkage
efficiency proposed by Kao (2019).
References
Galagedera, D. U. A., Roshdi, I., Fukuyama, H., & Zhu, J. (2018). A new network DEA model for
mutual fund performance appraisal: An application to U.S. equity mutual fund. Omega, 77,
168–179.
Kao, C. (2019). Inefficiency identification for closed series production systems. European Journal
Kao, C. (2020). Decomposition of slacks-based efficiency measures in network data envelopment
Lozano, S., & Khezri, S. (2021). Network DEA smallest improvement approach. Omega, 98,
102140. https://doi.org/10.1016/j.omega.2019.102140
Zhao, T. Y., Xie, J. H., Chen, Y., & Liang, L. (2022). Coordination efficiency in two-stage
network DEA: Application to a supplier-manufacturer sustainable supply chain. International
Journal of Logistics-Research and Applications, 25, 656–677.
Chapter 17
External and Internal Evaluations
There are two concepts used for measuring efficiencies in the network DEA: external
evaluation and internal evaluation. The former concept assesses efficiency from the
viewpoint of someone outside the organization, whereas the latter does so from the
viewpoint of the managers working inside the organization. In the case of the
external evaluation of the efficiency of a system such as a manufacturing plant, the
evaluator makes observations from a vantage point outside of the plant, seeing only
the exogenous inputs entering the plant and the exogenous outputs leaving it. The
evaluator has no direct knowledge of what has happened inside the plant. The
efficiency of the system in this case is defined as the ratio of the aggregate exogenous
output to the aggregate exogenous input, and on the basis of this ratio the perfor-
mance of different plants can be compared.
On the other hand, the assessment of the efficiency of the system can also be
conducted from within the organization. In this case, the internal evaluation is
performed by the general manager of the plant who is concerned with the operation
not only of the entire plant but also of its component divisions. The general manager
is satisfied only when all these divisions are operating efficiently. Therefore, the
efficiency of the system as a whole is considered as the aggregation of the efficien-
cies of all divisions, with all inputs and outputs occurring at all stages inside the
system taken into account.
In either evaluation, the efficiency of the system and its divisions can be measured
at the same time. Studies focusing on the external evaluation endeavor to decompose
the system efficiency into division efficiencies in order to determine what relation-
ship exists between the system and division efficiencies. This approach allows for
the identification of the divisions that are primarily responsible for the unsatisfactory
performance of the system. Making improvements to these divisions increases the
efficiency of the system as a whole in the most effective manner. On the other hand,
when the evaluation is carried out internally, the system efficiency is considered as
the aggregation of the efficiencies of all divisions, with those performing most
unsatisfactorily having the strongest impact on the efficiency of the system as a
https://doi.org/10.1007/978-3-031-27593-7_17
448 17 External and Internal Evaluations
whole. Studies on this type of evaluation tend to emphasize the mathematical

formulas of the aggregation so that the resulting model is easy to solve. These two
evaluative approaches have given rise to different measurements of the efficiency of
network systems, with the purpose of the measurement determining which of them is
the most appropriate.
The discussion presented in this chapter examines these two types of efficiency
evaluation. Section 17.1 discusses external evaluation and Sect. 17.2 examines
internal evaluation. For each type of evaluation, both the radial measure and the
slacks-based measure (SBM) of efficiency are considered. Finally, Sect. 17.3 pre-
sents relevant studies and applications related to this topic that have appeared in the
literature.
17.1 External Evaluation
In the context of carrying out an external evaluation, the efficiency of a network

system is defined as the ratio of the aggregate exogenous output to the aggregate
exogenous input. The objective of this assessment is to decompose the system
efficiency into division efficiencies so that the divisions that affect the performance
of the system the most can be identified.
Consider a general network system composed of p divisions which is structured
as shown in Fig. 17.1 and expressed using the notation introduced in Chap. 16. Each
ðk Þ
Division k of DMU j, j = 1,. . ., n, applies the exogenous inputs X ij , i 2 I(k), which
ðk Þ
have been supplied from outside the system, and the endogenous inputs Z fj , f 2
M(k), which have been produced by other divisions of the same system, to the
ðk Þ
production of both the exogenous outputs Y rj , r 2 O(k), which are to be sent out
ðk Þ
of the system, and the endogenous outputs Z gj , g 2 N(k), which are to be used by
other divisions in their own production process., where I(k), O(k), M(k), and N(k)
represent the index sets of the exogenous inputs and outputs as well as the
Fig. 17.1 Structure of the X i( k ) , i ∈ I ( k )

general network system
Z (f k ) Z g( k )
k
f ∈ M (k ) g ∈ N (k )
Yr( k ) , r ∈ O( k )
17.1 External Evaluation 449
endogenous inputs and outputs which are consumed and produced by a given
ðk Þ
Division k, respectively. The endogenous inputs Z f and endogenous outputs Z ðgkÞ
are intermediate products that are consumed and produced completely inside the
system and as such are invisible to outsiders. For the sake of simplicity, each
intermediate product is assumed to be produced by only one division and used by
only one division. It is further assumed that the system produces a total of
h intermediate products. We thus have [pk = 1 M ðkÞ = [pk = 1 N ðkÞ = {1, 2,. . ., h}. If
the exogenous inputs and outputs are also held to be exclusive to each division, with
totals of m inputs and s outputs, then we have [pk = 1 I ðkÞ = {1, 2,. . ., m} and [pk = 1 OðkÞ
= {1, 2,. . ., s}. However, the idea discussed in the following sections can be
extended to more general cases.
17.1.1 Radial Efficiency
When the system efficiency is defined as the ratio of the aggregate exogenous output
to the aggregate exogenous input, the radial efficiency of a DMU operating under
constant returns to scale can be calculated by means of the following model:
p ðkÞ
k=1 r2OðkÞ ur Y r0
E S0 = max: p ðkÞ
k=1 i2I ðkÞ vi X i0
ðkÞ ðkÞ
ðkÞ ðkÞ
r2O g2N
ðkÞ ðkÞ
vi X ij þ wf Z f j ≤ 0, j = 1, . . . , n, k = 1, . . . , p
ðkÞ ðkÞ
i2I f 2M
ur , vi , wg ≥ 0, r = 1, . . . , s, i = 1, . . . , m, g = 1, . . . , h:
ð17:1Þ
Recall from Chap. 9 that this model is a non-cooperative model, the intermediate
products produced by the supply division only need to be greater than or equal to
what is required by the demand division. In the case of a cooperative model where
the intermediate products produced by the supply division must be equal to what is
needed by the demand division, the multipliers wg associated with the intermediate
products are allowed to be negative. Since a positive or negative value of wg does not
affect the derivation of the relationship between the system and division efficiencies,
the non-cooperative Model (17.1) will be used in the following discussion.
After a set of optimal solutions (u*, v*, w*) for Model (17.1) has been obtained,
the system and division efficiencies can be calculated as:
p ðk Þ
k=1 r2OðkÞ ur Y r0
ES0 = p ðk Þ
k=1 i2I ðkÞ vi X i0
ð17:2Þ
ðk Þ ðk Þ
ðk Þ r2OðkÞ ur Y r0 þ g2N ðkÞ wg Z g0
E0 = , k = 1, . . . , p:
ðk Þ ðk Þ
i2I ðkÞ vi X i0 þ f 2M ðkÞ wf Z f 0
In a basic series system with a number of divisions connected in a series (see

Chap. 11), only the first division consumes exogenous inputs and only the last
division produces exogenous outputs, with all intermediate divisions consuming
only the endogenous inputs produced by the preceding division and producing
endogenous outputs for only the succeeding division to use. The efficiency of such
a system can be decomposed into the product of the efficiencies of all its component
divisions. In contrast, the component divisions of a parallel system operate indepen-
dently and there is no intermediate product connecting any two divisions. The
efficiency of this type of system can be decomposed into a weighted average of
the division efficiencies (see Chap. 12). Finally, by introducing dummy divisions
into the system which transmit factors, any network system without feedback can be
converted into a series of subsystems each of which is composed of a number of
divisions connected in parallel.
Suppose that the converted system is composed of t subsystems and the dth
subsystem is composed of | Sd | divisions connected in parallel, with Sd the set that
ðd Þ
contains the indexes of the divisions in Subsystem d. Denote e0 as the efficiency of
ðd Þ
Subsystem d and ω as the weight associated with Division k when expressing e0
(k)
as a weighted average of the efficiencies of its component divisions. We have:
t
ðd Þ
E S0 = e0 , ð17:3Þ
d=1
where
ðd Þ ðk Þ
e0 = ωðkÞ E 0 , with ωð k Þ = 1 and ωðkÞ ≥ 0, k 2 Sd :
k2Sd k2Sd
This decomposition is made possible by converting the original system into

several parallel subsystems connected in a series. Moreover, it is necessary to
determine the weights ω(k) to be applied in the decomposition. If only a rough idea
of the decomposition is desired, one can use the bounds of the decomposition. The
ðk Þ ðk Þ
division efficiency, as defined in Eq. (17.2), is E0 = [ r2OðkÞ ur Y r0 +
g2N ðkÞ wg Z g0 ]/[ i2I ðkÞ vi X i0 + f 2M ðkÞ wf Z f 0 ]. Let the weight ω associated
with Division k be expressed as the aggregate input consumed by this division in
that consumed by all divisions in the same parallel subsystem. That is, ωðkÞ =
ðk Þ ðk Þ ðkÞ ðk Þ
[ i2I ðkÞ vi X i0 + f 2M ðkÞ wf Z f 0 ] / pk = 1 ½ i2I ðkÞ vi X i0 + f 2M ðkÞ wf Z f 0 ]. It is obvious
ðk Þ
that p
k = 1ω = 1 and ωðkÞ ≥ 0, k = 1,. . ., p. The weighted average of the
efficiencies of all p divisions can be expressed as:
p
ðk Þ
ωðkÞ E 0
k=1
i2I ðkÞ vi X i0 þ r2OðkÞ ur Y r0 þ
p
f 2M ðkÞ wf Z f 0 g2N ðkÞ wg Z g0
= ×
i2I ðkÞ vi X i0 þ
p
k=1 k=1 i2I ðkÞ vi X i0 þ f 2M ðkÞ wf Z f 0 f 2M ðkÞ wf Z f 0
p ðk Þ ðk Þ
k=1 r2OðkÞ ur Y r0 þ g2N ðkÞ wg Z g0
= :
p ðk Þ ðk Þ
k=1 i2I ðkÞ vi X i0 þ f 2M ðkÞ wf Z f 0
ðk Þ ðk Þ
Note that pk = 1 g2N ðkÞ wg Z g0 = pk = 1 f 2M ðkÞ wf Z f 0 , which means that the
aggregation of all the intermediate products produced in the system is equal to the
aggregation of all the intermediate products consumed in the system. Denote ρ as the
ratio of this value to the aggregation of all exogenous inputs in the system. We obtain
the following expression by dividing both the numerator and the denominator by the
ðk Þ
aggregate exogenous input pk = 1 i2I ðkÞ vi X i0 :
p ðk Þ p ðk Þ
p
ðk Þ ES0 þ k=1 g2N ðkÞ wg Z g0 = k=1 i2I ðkÞ vi X i0 E S0 þ ρ
ωðkÞ E 0 = = ≥ E S0 :
1þ p ðk Þ p ðk Þ 1þρ
k=1 k=1 f 2M ðkÞ wf Z f 0 = k=1 i2I ðkÞ vi X i0
The network DEA model with the objective of finding the multipliers (u, v, w)
that produce the maximum weighted average of the division efficiencies,
p ðk Þ ðk Þ
k = 1 ω E 0 , is:
p ðk Þ p ðk Þ
k=1 r2OðkÞ ur Y r0 þ k=1 g2N ðkÞ wg Z g0
σ = max p ðk Þ p ðk Þ
k=1 i2I ðkÞ vi X i0 þ k=1 f 2M ðkÞ wf Z f 0
ðk Þ ðk Þ
r2OðkÞ g2N ðkÞ
ðk Þ ðk Þ
i2I ðkÞ f 2M ðkÞ
ur , vi , wg ≥ 0, r = 1, . . . , s, i = 1, . . . , m, g = 1, . . . , h
ð17:4Þ
At optimality, we have σ = (ES0 +ρ)/(1 + ρ) ≥ ES0 . In other words, the system

efficiency E S0 is bounded from above by σ, the weighted average of the division
efficiencies. If the weights ωðkÞ are similar for all divisions, then the system
efficiency is less than or equal to the average of the division efficiencies.
In Eq. (17.3), the subsystem efficiency is a weighted average of the efficiencies of
ðd Þ ðk Þ
the component divisions, so e0 = k2Sd ωðkÞ E 0 . Given that the weighted arith-
metic average is greater than or equal to the weighted geometric average, we have
the following relationships:
t t t ðk Þ p ðk Þ
ðd Þ ðk Þ ðk Þ ω ðk Þ ω
E S0 = e0 = ωð k Þ E 0 ≥ E0 = E0 :
d=1 d=1 k2Sd d=1 k2Sd k=1
The last equality follows because dummy divisions have an efficiency score of
one, which can be ignored in the multiplication, and only real divisions are left.
Denote π as the multiplicative aggregation of the division efficiencies of the form
ðk Þ
p ðk Þ ω
π= k=1 E0 . We then have the following relationship:
p ωðk Þ
ðk Þ
E S0 ≥ E0 = π:
k=1
This relationship shows that the system efficiency is bounded from below by the
weighted geometric average of the division efficiencies. If the weights ω(k) are
similar for all divisions, then the system efficiency is greater than or equal to the
geometric average of the division efficiencies.
Together with the upper bound obtained previously, we conclude that the effi-
ciency of a network system is bounded from above by the arithmetic mean of the
division efficiencies and from below by their geometric mean:
p ðk Þ p
ðk Þ ω ðk Þ
π= E0 ≤ E S0 ≤ ωðkÞ E 0 = σ: ð17:5Þ
k=1 k=1
Note that the weights ω(k) used in calculating the weighted geometric average are
different from the weights ωðkÞ used in calculating the weighted arithmetic average.
The bounds suggest that if the system efficiency is close to the upper bound σ, then
the structure of the system more closely resembles a parallel system. In this case,
divisions with larger weights ωðkÞ in the additive aggregation functions have a
stronger effect on the performance of the system as a whole, and so they are the
divisions that the decision-maker should pay more attention to. On the other hand, if
the system efficiency is close to the lower bound π, then the structure of the system
more closely resembles a series system. In this case, divisions with larger weights
ω(k) in the multiplicative aggregation functions have a stronger effect on the perfor-
mance of the system as a whole, and so they are the divisions that must be prioritized
in order to improve the efficiency of the system the most.
The exact relationship between the system efficiency E S0 and the division effi-
ðk Þ
ciencies E 0 can be calculated by converting the structure of the system into a series
of subsystems each of which is composed of several divisions connected in parallel.
This approach to efficiency decomposition has been discussed in the preceding
chapters on network systems with various structures, and so it will not be
repeated here.
17.1.2 SBM Efficiency
The decomposition of the system efficiency discussed in the preceding subsection is

based on radial efficiencies. It can also be discussed in terms of the SBM efficiency
which functions under both constant and variable returns to scale, in contrast to
radial efficiencies which are applicable only with constant returns to scale. It should
be noted, however, that while the decomposition of a radial efficiency is identical in
both cooperative and non-cooperative cases, the decomposition of an SBM effi-
ciency differs according to the type of case being considered.
The SBM model for measuring the efficiency of the network system shown in
Fig. 17.1 for the cooperative case is:
1 p ðkÞ - ðkÞ
1- k=1
s
i2I ðkÞ i
=X i0
ρ = min:
S m
1 p ðkÞ
1þ k=1
sðkÞþ =Y r0
r2OðkÞ r
s
n
s:t: λj X i j þ si = X i0 , i 2 I ðkÞ , k = 1, . . . , p
j=1
n
ðkÞ ðkÞ ðkÞ
λj Y r j - srðkÞþ = Y r0 , r 2 OðkÞ , k = 1, . . . , p
j=1
n n
λj s Z g js = λj Zg j , g = 1, . . . , h
j=1 j=1
n
ðkÞ
λj = 1, k = 1, . . . , p
j=1
λj , si , sr ≥ 0, j = 1, . . . ,n, k = 1, . . . ,p, i = 1, . . . ,m, r = 1, . . . , s:
ð17:6Þ
where gs and gd denote the divisions that produce and consume the intermediate
product Zg, respectively. This model was introduced in Chap. 16 as Model (16.16). If
the equality constraint nj = 1 λj s Z gj s = nj = 1 λj d Z gj d is changed to the inequality
n ðgs Þ ðgs Þ n ðgd Þ ðgd Þ
constraint j = 1 λj Z gj ≥ j = 1 λj Z gj , then it becomes a non-cooperative
model.
To measure the efficiency of Division k, we suppose that
n
ðkÞ ðk Þ ðk Þ - ðk Þ
λj Z fj þ t f = Zf 0 , f 2 M ðk Þ , k = 1, . . . , p
j=1
n
ðkÞ ðk Þ ðk Þ
λj Z gj - t ðgkÞþ = Z g0 , g 2 N ðkÞ , k = 1, . . . , p:
j=1
ðk Þ -
Note that the slack variables t f and t ðgkÞþ are unrestricted in sign. The efficiency
of Division k and the linkage efficiency Zg are calculated by means of:
1 ðkÞ - ðkÞ ðkÞ - ðkÞ

1- ðkÞ ðkÞ ðkÞ si =X i0 þ 2M ðkÞ t f =Z f 0
jI jþjM j i2I f
ρðkÞ = , k = 1, . . . , p
1 ðkÞþ ðkÞ ðkÞþ ðkÞ
s =Y þ t
g2N ðkÞ g
=Z
r0 g0
1 ðgd Þ - ðg Þ
1 - ðgÞ ðgÞ t g =Z g0d
jL j g2L
πðgs ,gd Þ = , g = 1, . . . , h,
1 ðgs Þþ ðgs Þ
1 þ ðgÞ g2LðgÞ g
t =Z g0
jL j
ð17:7Þ
where L(g) is the index set of the intermediate products Zg produced by Division gs
and used by Division gd, and | I(k) |, | O(k) |, | M(k) |, | N(k) |, and | L(g) | denote the
numbers of elements in the corresponding sets.
As discussed in Chap. 16, the efficiency of the system can be decomposed into the
product of the division efficiencies adjusted by the product of the linkage efficiencies
for series systems, and it can be decomposed into a weighted average of the division
efficiencies for parallel systems. By converting the structure of the network system
into a series of components each of which is a subsystem of divisions connected in
parallel, the relationship between the system efficiency and the division efficiencies
can be obtained. The relevant derivations are discussed in detail in Chap. 16, and so
they will not be repeated here.
17.2 Internal Evaluation
When an internal evaluation is carried out, the efficiency of each division of a system
is measured first. The system efficiency is then defined as the aggregation of the
division efficiencies. A division that is considered more important is assigned a
larger weight in the aggregation function. Whereas carrying out an external evalu-
ation involves decomposing the system efficiency with the goal of identifying the
divisions that act most prominently in the unsatisfactory performance of the system,
17.2 Internal Evaluation 455
performing an internal evaluation aims to determine the aggregation of the division

efficiencies that best reflects the impact of the unsatisfactory performance of the
divisions on the efficiency of the system as a whole. In short, the point of aggregating
efficiencies is to define the system efficiency as a function of the division efficien-
ð1Þ ð2Þ ðpÞ
cies: ES0 = F(E 0 , E0 ,. . ., E0 ). The corresponding model is designed to determine
the multipliers, in the radial model, or intensity coefficients, in the SBM model, that
produce the appropriate aggregated division efficiencies. Theoretically, the function
for aggregating the division efficiencies can take any mathematical formulation. The
division efficiency can be measured under either constant or variable returns to scale,
and the model can be either cooperative or non-cooperative.
As was done in Sect. 17.1, the following discussion is divided into subsections
dealing with radial efficiency and SBM efficiency.
17.2.1 Radial Efficiency
For calculating radial efficiencies under the more general case of variable returns to
scale, the cooperative input model for measuring the system efficiency of a general
network system is:
ð1Þ ð2Þ ðpÞ

E0S = max: F E0 , E0 ; , ; . . . ; , E0
p p
ðkÞ ðkÞ
s:t: ur Y r j - u0 - vi X i j ≤ 0, j = 1, . . . , n
k = 1 r2OðkÞ k = 1 i2I ðkÞ
ðkÞ ðkÞ ðkÞ

ur Y r j þ wg Z g j ; - u0 -
ðkÞ ðkÞ
r2O g2N
ðkÞ ðkÞ
vi X i j þ wf Z f j ≤ 0, k = 1, . . . ,p, j = 1, . . . , n
i2I ðkÞ f 2M ðkÞ
ur , vi ≥ ε, r = 1, . . . ,s, i = 1, . . . , m
wg unrestricted in sign, g = 1, . . . , h,
ð17:8Þ
ðk Þ
where u0 = pk = 1 u0 . The first and second sets of constraints are associated with the
system and the p divisions, respectively. Since each constraint in the first set is the
sum of the p constraints in the second set corresponding to the same DMU, the first
set of constraints is redundant and so can be eliminated. For cases of constant returns
ðk Þ
to scale, u0 and u0 are deleted, and the multipliers wg are restricted to be nonneg-
ative in the context of the non-cooperative model. These modifications have no
effect on the aggregation of the division efficiencies.
At optimality, the division efficiencies are calculated as:
ðk Þ r2OðkÞ ur Y r0 þ g2N ðkÞ wg Z g0 - u0
E0 = ðk Þ ðk Þ
, k = 1, . . . , p,
i2I ðk Þ vi X i0 þ f 2M ðk Þ wf Z f 0
which is essentially identical to Eq. (17.2), except that the efficiency in Eq. (17.2) is
ðk Þ
calculated under constant returns to scale without the u0 term.
Two mathematical expressions of the function F have been used when aggregat-
ing the division efficiencies to constitute the system efficiency: the multiplicative
and additive forms. As shown above, the efficiency of Division k is calculated as:
ðk Þ ðk Þ ðk Þ ðk Þ ðk Þ ðk Þ
E 0 = [ r2OðkÞ ur Y r0 + g2N ðkÞ wg Z g0 - u0 ] / [ i2I ðkÞ vi X i0 + f 2M ðkÞ wf Z f 0 ]. The
system efficiency of the multiplicative form is the product of the division efficien-
ðk Þ
cies: ES0 = pk = 1 E 0 . This function has only been used for two-division systems. For
example, when the first division does not produce final outputs, the objective
function is: ES0 = max [( g2N ð1Þ wg Z g0 - u0 )/ i2I ð1Þ vi X i0 ] × [( r2Oð2Þ ur Y r0 -
u0 )/( i2I ð2Þ vi X i0 + f 2M ð2Þ wf Z f 0 )]. Note that g2N ð1Þ wg Z g0 = f 2M ð2Þ wf Z f 0 in
this case. A weakness of this model is that the objective function is nonlinear, which
requires powerful computer software packages to solve.
For the additive function, the system efficiency is defined as a weighted average
ðk Þ
of the division efficiencies: E S0 = pk = 1 ωðkÞ E0 , where ω(k) ≥ 0, with pk = 1 ωðkÞ =
1, representing the importance of Division k to the efficiency of the system. The
weights can be either pre-specified externally by the decision-maker or reflected
internally from the observations. If the weights are pre-specified constants, for
example, 1/p for each division, then the corresponding objective function is:
p
ðk Þ
p
1 r2OðkÞ ur Y r0 þ g2N ðkÞ wg Z g0 - u0
ωð k Þ E 0 = :
p ðk Þ þ ðk Þ
k=1 k=1 i2I ðkÞ vi X i0 f 2M ðkÞ wf Z f 0
Again, it is nonlinear which makes the problem difficult to solve.

For reflected weights, the weight associated with Division k is usually defined as
the ratio of the aggregate input of this division to that of all divisions (Cook et al.,
2010), i.e. ω(k) = [ i2I ðkÞ vi X i0 + f 2M ðkÞ wf Z f 0 ]/
p
k=1 i2I ðkÞ vi X i0 +
ðk Þ
f 2M ðkÞ wf Z f 0 ]. With this set of weights, the system efficiency becomes:
p
ðk Þ
E S0 ¼ ωðkÞ E 0
k¼1
ðk Þ ðk Þ
p
i2I ðkÞ vi X i0 þ f 2M ðkÞ wf Z f 0
¼
p ðk Þ þ ðk Þ
k¼1 k¼1 i2I ðkÞ vi X i0 f 2M ðkÞ wf Z f 0
ð17:9Þ
r2OðkÞ ur Y r0 þ g2N ðkÞ wg Z g0 - u0
×
ðk Þ þ ðk Þ
i2I ðkÞ vi X i0 f 2M ðkÞ wf Z f 0
p ðk Þ ðk Þ ðk Þ
k¼1 r2OðkÞ ur Y r0 þ g2N ðkÞ wg Z g0 - u0
¼
p ðk Þ þ ðk Þ
k¼1 i2I ðkÞ vi X i0 f 2M ðkÞ wf Z f 0
The system efficiency is simplified by reducing the number of fractionals from

p fractionals to one fractional. The objective function can then be linearized by
applying a variable substitution technique (Charnes & Cooper, 1962) to make the
problem easier to solve.
The efficiency aggregation presented in Eq. (17.9) is discussed from the input
side. However, it can also be discussed from the output side. Conventionally, the
objective function of the output model is expressed as the minimization of the
ðk Þ
inverse of the efficiency: 1/ E S0 = pk = 1 ωðkÞ =E 0 , which suggests that the system
efficiency is a weighted harmonic average of the division efficiencies. Wang and
Chin (2010) defined the weight associated with a division as the ratio of the
aggregate output of this division to that of the p divisions. Thus, they were able to
derive the same objective function of one fractional under constant returns to scale,
yet in an inverse form.
Recall that the overall efficiency of a parallel system under constant returns to
scale can be decomposed into a weighted average of the division efficiencies. Since
the weighted average is in the form of an additive function, the system efficiency
calculated in the efficiency decomposition approach is the same as that obtained
from the efficiency aggregation approach.
This method of efficiency aggregation can be applied to network systems of any
structure, even those with feedback or cycles. Once a mathematical formulation for
aggregation is determined, the system efficiency can then be expressed as the
objective function presented in Model (17.8), which is easy to solve.
17.2.2 SBM Efficiency
The SBM model is expressed in envelopment form, with the same constraints as
those shown in Model (17.6). The efficiency of Division k, as shown in Eq. (17.7) is:
ðk Þ - ðk Þ ðk Þ - ðk Þ
1- 1
jI ðkÞ jþjM ðkÞ j i2I ðkÞ si =X i0 þ f 2M ðkÞ t f =Z f 0
ρðkÞ = , k = 1, . . . , p:
ðk Þþ ðk Þ ðk Þþ ðk Þ
1 þ jOðkÞ jþjN
1
ðk Þ
j r2O ðk Þ sr =Y r0 þ g2N ðk Þ tg =Z g0
In the internal evaluation, the efficiency of the system is defined as a function of

the division efficiencies: ρS = F(ρ(1),. . ., ρ( p)).
Recall that the SBM efficiency is the product of the arithmetic average of the
input efficiencies and the harmonic average of the output efficiencies. Suppose the
importance of Division k is ω(k). In order to maintain linearity, Tone and Tsutsui
(2009) defined the efficiency of the system as the product of the arithmetic average
of the input efficiencies weighted by ω(k) and the harmonic average of the output
efficiencies weighted by ω(k). The complete model is:
p ðk Þ 1 ðkÞ- ðk Þ ðkÞ- ðk Þ
k¼1 ω 1- s =X i0 þ t =Z f 0
j I j þ j M ðkÞ j
ðk Þ i2I ðkÞ i f 2M ðkÞ f
ρS ¼ min:
1 ðk Þ ðk Þ
p
k¼1 ω
ðk Þ 1þ ðkÞ sðkÞþ =Y r0 þ
r2OðkÞ r
t ðkÞþ =Z g0
g2N ðkÞ g
j O j þ j N ðk Þ j
n
ðk Þ ðk Þ ðkÞ- ðk Þ
s:t: λ j X ij þsi ¼ X i0 , i 2 I ðkÞ , k ¼ 1, ...,p
j¼1
n
λ j Y rj -sðrkÞþ ¼ Y r0 , r 2 OðkÞ , k ¼ 1, ...,p
j¼1
n n
λ j s Z gj s ¼ λ j d Z gj d , g ¼ 1, ...,h
j¼1 j¼1
n
ðk Þ
λ j ¼ 1, k ¼ 1, ...,p
j¼1
ðk Þ ðkÞ-
λj , si , sðrkÞþ ≥0, j ¼ 1,...,n, k ¼ 1,...,p, i ¼ 1,...,m, r ¼ 1, ...,s:
ð17:10Þ
In this formulation, the intermediate products have been included in the expres-
sion of the efficiency of the system. This model can be linearized by applying a
variable substitution technique proposed in Charnes and Cooper (1962).
The weight ω(k) in Model (17.10) is a pre-specified weight for Division k,
representing its importance to the system. It must have a positive value and add up
to 1 over all p divisions: ω(k) ≥ 0 and qp = 1 ωðkÞ = 1. Tone and Tsutsui (2009) also
defined the numerator and denominator (in reciprocal form) of the objective function
in Model (17.10) as the input- and output-oriented efficiencies of the system,
respectively. Under these definitions, the input and output efficiencies of the system
become weighted arithmetic mean and the weighted harmonic mean, respectively, of
the efficiencies of the component processes. The weights ω(k) can also be reflected
from the observations. In this case the contribution of Division k to the system
efficiency is determined from the observations. Different from Model (17.10) where
the same weight is applied to the same division for every DMU, here the weight
calculated for the same division of different DMUs may be different.
As it is not necessary to specify the weight for each division beforehand, ω(k) in
Model (17.10) can be determined afterward. Specifically, one may define the weight
ω(k) as the ratio of the output efficiency of Division k (in reciprocal form) to the sum
of the output efficiencies of all divisions:
ðkÞþ ðkÞ ðkÞþ ðkÞ

1 þ jOðkÞ jþjN
1
ðkÞ
j r2OðkÞ
sr =Y r0 þ g2N ðkÞ
tg =Z g0
ðkÞ
ω = p :
1 ðkÞ ðkÞ
1þ sðkÞþ =Y r0
r2OðkÞ r
þ t ðkÞþ =Z g0
g2N ðkÞ g
k=1 j OðkÞ j þ j N ðkÞ j
The weighted average of the efficiencies of all p divisions then becomes:
p
ρS = ωðkÞ ρðkÞ
k=1
1 ðkÞ ðkÞ
1þ r2OðkÞ r
sðkÞþ =Y r0 þ g2N ðkÞ g
t ðkÞþ =Z g0
p
j OðkÞ j þ j N ðkÞ j
= p ×
k=1 1 ðkÞ ðkÞ
1 þ ðkÞ sðkÞþ =Y r0 þ
r2OðkÞ r g2N ðkÞ g
t ðkÞþ =Z g0
k=1 j O j þ j N ðkÞ j
1 - ðkÞ ðkÞ ðkÞ si =X i0 þ 2M ðkÞ t f =Z f 0
jI jþjM j i2I f
1 ðkÞ ðkÞ
sðkÞþ =Y r0 þ g2N ðkÞ g
t ðkÞþ =Z g0
p
1 - ðkÞ ðkÞ
s
i2I ðkÞ i
=X i0 þ f 2M ðkÞ f
t =Z f 0
k=1 j I j þ j M j
= p :
1 ðkÞþ ðkÞ ðkÞþ ðkÞ
s =Y þ g2N ðkÞ g
t =Z
r0 g0
k=1
The model for calculating the system efficiency can be formulated as (Kao,
2014b):
p
1- s =X i0 þ t =Z f 0
k=1 j I j þ j M ðkÞ j
ðkÞ i2I ðkÞ i f 2M ðkÞ f
ρS = min: p
1 ðkÞ ðkÞ
1 þ ðkÞ sðkÞþ =Y r0
r2OðkÞ r
þ t ðkÞþ =Z g0
g2N ðkÞ g
k=1 j O j þ j N ðkÞ j
n
s:t: λj X ij þ si = X i0 , i 2 I ðkÞ , k = 1, . . . , p
j=1
n
λj Y rj - sr = Y r0 , r 2 OðkÞ , k = 1, . . . , p
j=1
n n
λj s Z gj s = λj Z gj , g = 1, . . . , h
j=1 j=1
n
ðkÞ
λj = 1, k = 1, . . . , p
j=1
λ j , si , sr ≥ 0, j = 1, . . . , n, k = 1, . . . , p, i = 1, . . . , m, r = 1, . . . , s:
ð17:11Þ
Tone and Tsutsui (2014) also extended their network SBM model to a dynamic
environment, where the network system repeats for a number of periods and two
consecutive periods are connected by carryovers. With a set of pre-specified weights
for each division in the network and another set of weights for each period, the
overall efficiency of a DMU over T periods can be defined in a same manner. The
concept of reflected weights can also be used to express the overall efficiency as a
weighted average of the division efficiencies of every period. In this case, the
division and period weights do not need to be specified beforehand. They can instead
be derived after the overall efficiency has been calculated by setting the division
weight as that in the static environment and the period weight as the ratio of the
output-oriented efficiency score for period t to the sum of the output-oriented
efficiency scores for all T periods. The weights corresponding to either a division
or a period reflect the importance of this division or period to the overall efficiency as
perceived by the DMU being evaluated.
From a methodological point of view, decomposing the system efficiency into

division efficiencies while carrying out an external evaluation is comparatively
more difficult than aggregating the division efficiencies to form the system efficiency
by means of an internal evaluation. From the practical point of view, however, there
is no difference between these ideas.
The idea of efficiency decomposition in the context of calculating radial efficien-
cies was first introduced in Kao and Hwang (2008) for two-stage systems. It was
then extended to general network systems in Kao (2009). This idea works only under
constant returns to scale. The bounds for the system efficiency were derived in Kao
(2018), also under constant returns to scale. In terms of the application of this idea,
Hsieh and Lin (2010) measured the efficiency of 57 tourist hotels in Taiwan where
the system consisted of two principal functions, catering and accommodation,
operated independently and each function was divided into two stages of production
and consumption. Accordingly, the efficiency of the system was decomposed into
the efficiencies of four divisions. Other relevant work has examined the service and
sales efficiencies of 20 branches of a Canadian bank (Cook et al., 2000); the
performance of 39 branches of a commercial bank in Iran, with the functions of
deposits, sales, and services (Jahanshahloo et al., 2004); the efficiency of 10 steel
plants, each of which comprised four production components (Cook & Green,
2004); the cost efficiency of treating solid waste in 293 municipalities in Flanders,
Belgium (Rogge & Jaeger, 2012); and the efficiency of the drinking water and
wastewater services of 45 water utilities in Portugal (Da Cruz et al., 2013). In
another study, Kao (2015) examined a hierarchical system composed of a number
of divisions each of which consisted of several subordinate subdivisions. This
system was shown to be equivalent to a basic parallel system in that it was possible
to decompose the system efficiency into a weighted average of the efficiencies of the
divisions at the bottom of the hierarchical structure. Finally, Kao (2013, 2014a) also
identified a relationship between the system and division efficiencies for a general
series system, where all divisions were allowed to consume exogenous inputs and
produce exogenous outputs.
With regard to efficiency aggregation, Chen et al. (2006), Zha and Liang (2010),
and Li et al. (2012) used a cooperative game approach to study a supply chain
problem where the system efficiency was defined as the product of the two divisions.
In addition, Liang et al. (2006) and Bichou (2011) assigned equal weights to the two
division efficiencies to constitute the system efficiency while assessing the efficiency
of supply chains. In the case of reflected weights, Chen et al. (2009) defined the
weight of a division to be the proportion of the aggregate input consumed by this
division in the total aggregate input consumed by all divisions for a basic two-stage
system in which the first division does not produce exogenous outputs and the
second division does not consume exogenous inputs. This relationship was used
by Guan and Chen (2012) to measure the R&D and commercialization efficiencies
of 22 OECD countries, and it has been generalized for use with systems consisting of
multiple stages (Amirteimoori & Kordrostami, 2005; Amirteimoori & Shafiei, 2006
and Kordrostami & Amirteimoori, 2005) and with network systems of all types of
structures (Cook et al., 2010).
The concept of efficiencies aggregation can be applied to network systems of any
structure, even those with feedback or cycles (Cheng & Gao, 2015; Kawaguchi et al.,
2014; Liang et al., 2011; Shafiee et al., 2014 and Yang et al., 2014). Once a
mathematical formulation for aggregation is determined, the system efficiency is
then defined and the corresponding model is constructed for calculating efficiencies.
References
Amirteimoori, A., & Kordrostami, S. (2005). DEA-like models for multi-component performance
Amirteimoori, A., & Shafiei, M. (2006). Measuring the efficiency of interdependent decision
making sub-units in DEA. Applied Mathematics and Computation, 173, 847–855.
Bichou, K. (2011). A two-stage supply chain DEA model for measuring container-terminal
efficiency. International Journal of Shipping and Transport Logistics, 3, 6–26.
Chen, Y., Cook, W. D., Li, N., & Zhu, J. (2009). Additive efficiency decomposition in two-stage
Chen, Y., Liang, L., & Yang, F. (2006). A DEA game model approach to supply chain efficiency.
Cheng, Y., & Gao, H. L. (2015). Matrix-type network DEA model with its application based on
input-output tables. Mathematical Problems in Engineering, 2015, 1–8. https://doi.org/10.1155/
2015/505941
Cook, W. D., & Green, R. H. (2004). Multicomponent efficiency measurement and core business
identification in multiplant firms: A DEA model. European Journal of Operational Research,
157, 540–551.
Cook, W. D., Hababou, M., & Tuenter, H. J. H. (2000). Multicomponent efficiency measurement
and shared inputs in data envelopment analysis: An application to sales and service performance
in bank branches. Journal of Productivity Analysis, 14, 209–224.
Cook, W. D., Zhu, J., Bi, G. B., & Yang, F. (2010). Network DEA: Additive efficiency decompo-
sition. European Journal of Operational Research, 207, 1122–1129.
Da Cruz, N. F., Carvalho, P., & Marques, R. C. (2013). Disentangling the cost efficiency of jointly
provided water and wastewater services. Utilities Policy, 24, 70–77.
Guan, J. C., & Chen, K. H. (2012). Modeling the relative efficiency of national innovation systems.
Research Policy, 41, 102–115.
Hsieh, L. F., & Lin, L. H. (2010). A performance evaluation model for international tourist hotels in
Taiwan: An application of the relational network DEA. International Journal of Hospitality
Jahanshahloo, G. R., Amirteimoori, A. R., & Kordrostami, S. (2004). Multi-component perfor-
mance, progress and regress measurement and shared inputs and outputs in DEA for panel data:
An application in commercial bank branches. Applied Mathematics and Computation., 151,
1–16.
model. European Journal of Operational Research., 192, 949–962.
Kao, C. (2014a). Efficiency decomposition for general multi-stage systems in data envelopment
Kao, C. (2015). Efficiency measurement for hierarchical network systems. Omega, 51, 121–127.
Kao, C. (2018). Multiplicative aggregation of division efficiencies in network data envelopment
Research, 185, 418–429.
Kawaguchi, H., Tone, K., & Tsutsui, M. (2014). Estimation of the efficiency of Japanese hospitals
using a dynamic and network data envelopment analysis model. Health Care Management
Science, 17, 101–112.
References 463
Kordrostami, S., & Amirteimoori, A. (2005). Undesirable factors in multi-component performance

Li, Y., Chen, Y., Liang, L., & Xie, J. H. (2012). DEA models for extended two-stage network
structures. Omega, 40, 611–618.
Liang, L., Li, Z. Q., & Cook, W. D. (2011). Data envelopment analysis efficiency in two-stage
networks with feedback. IIE Transactions, 43, 309–322.
Liang, L., Yang, F., Cook, W. D., & Zhu, J. (2006). DEA models for supply chain efficiency
evaluation. Annals of Operations Research, 145, 35–49.
Rogge, N., & Jaeger, S. (2012). Evaluating the efficiency of municipalities in collecting and
processing municipal solid waste: A shared input DEA-model. Waste Management, 32,
1968–1978.
Shafiee, M., Lotfi, F. H., & Saleh, H. (2014). Supply chain performance evaluation with data
envelopment analysis and balanced scorecard approach. Applied Mathematical Modeling, 38,
5092–5112.
Tone, K., & Tsutsui, M. (2014). Dynamic DEA with network structure: A slacks-based measure
Wang, Y. M., & Chin, K. S. (2010). Some alternative DEA models for two-stage process. Expert
Systems with Applications, 37, 8799–8808.
Yang, F., Du, F., Liang, L., & Yang, Z. (2014). Forecasting the production abilities of recycling
systems: A DEA based approach. Journal of Applied Mathematics, 2014, 1–9. https://doi.org/
10.1155/2014/961468
Zha, Y., & Liang, L. (2010). Two-stage cooperation model with input freely distributed among the
stages. European Journal of Operational Research, 205, 332–338.
Chapter 18
Epilogue
Network data envelopment analysis is a relatively new subject, with a short history
of no more than twenty-five years since the term first appeared in 2000. However,
due to its ability to measure the performance of complex systems, it has been widely
applied to real world problems. As a result of addressing such problems, new ideas
have emerged which improve the existing methods, and yield better solutions. In the
last two decades dozens of models have been proposed, and new models are still
being developed. This raises the following two issues: which model to use when
faced with a specific problem, and where to go next. While it is impossible to give
exact and complete answers to these questions, this chapter provides some useful
directions to think about how to answer them. Before getting into more detail, we
first discuss the generality of some major classes of models.
18.1 Generality of Some Representative Models
Every model for measuring efficiency is based on the production possibility set
defined by the DMUs being evaluated, within which a point is selected as the
benchmark for a DMU to compare with. Consider the general network system
ðk Þ
where each Division k applies the exogenous inputs X i , i = 1, . . ., m, and the
ða,kÞ
intermediate inputs Z f , f = 1, . . ., h, produced by Division a, to produce the
intermediate outputs Z ðgk,bÞ , g = 1, . . ., h, for Division b to use, and the final outputs
Y ðrkÞ , r = 1, . . ., s, to send out of the system (referring to the structure shown in
Fig. 9.2). In the most general form, the constraints defining the production possibility
set, along with some parameters for measuring the efficiencies of the network
system, can be expressed as:
https://doi.org/10.1007/978-3-031-27593-7_18
466 18 Epilogue
n
λj X ij ≤ θi X i0 , 8i, k
j=1
n p ðkÞ p
≤ ^θf
ðkÞ ða,kÞ ða,kÞ
λj Zf j Zf 0 , 8f , k
j=1 a=1 a=1
n p p
ðkÞ ðk,bÞ ðk,bÞ
λj Z gj ^ gðkÞ
≥φ Z g0 , 8g, k
j=1 b=1 b=1
n
ð18:1Þ
λj Y rj ≥ φr Y r0 , 8r, k
j=1
ðkÞ
θi , ^θf ≤ 1,
ðkÞ
8i, f , k
^ gðkÞ , φrðkÞ
φ ≥ 1, 8g, r, k
ðkÞ
λj ≥ 0, 8j, k,
where θi , θf , ϕg , and ϕðrkÞ are distance parameters. Different restrictions on
these parameters result in different models. Here we discuss some typical ones. Note
that the production possibility set that the efficiency measures are based on is
considered under the assumption of constant returns to scale. For cases of variable
ðk Þ
returns to scale, the convexity constraint nj = 1 λj = 1 is added. Furthermore, if a
ðk Þ
free disposal hull is desired, then λj 2 f0, 1g is needed. However, these modifica-
tions do not affect the relationships to be derived below.
1. Slacks-based model
Replacing the distance parameters in Set (18.1) with slack variables via
θi = X i0 - si =X i0 , 8i, k
p p
θf = Z f 0 - sf = Zf 0 , 8f , k
a=1 a=1
ð18:2Þ
p p
ðk,bÞ ðk,bÞ
φðgkÞ = Z g0 þ sðgkÞþ = Z g0 , 8g, k
b=1 b=1
ðk Þ ðk Þ
φðrkÞ = Y r0 þ sðrkÞþ =Y r0 , 8r, k
obtains the feasible region of the slacks-based model. The feasible regions
corresponding to Sets (18.1) and (18.2) are equivalent.
2. Factor parameter model
In Set (18.1), if we require the distance parameters corresponding to the same
input factor of different divisions to be the same, and all other parameters to be
ignored, that is,
18.1 Generality of Some Representative Models 467
ðk Þ
θ i = θi , 8i, k
ðk Þ ð18:3Þ
θf = φðgkÞ = φðrkÞ = 1, 8f , g, r, k
then we have a constraint set which is the same as that of the input-oriented factor
parameter model. This set is clearly contained in Set (18.2).
ðk Þ
The output-oriented counterpart is obtained by letting ϕðrkÞ = ϕr and θi =
ðk Þ ðk Þ
θf = ϕg = 1: The resulting set is also contained in Set (18.2). In this respect, the
slacks-based model is more general than the factor parameter model.
3. Division parameter model
Suppose the parameters corresponding to the same division of different factors in
Set (18.1) are restricted to be the same, and all other parameters are discarded, that is,
ðk Þ
θi = θðkÞ , 8i, k
ðk Þ ð18:4Þ
θf = φðgkÞ = φðrkÞ = 1, 8f , g, r, k
We have the constraint set of the input-oriented division parameter model.

Obviously, this set is smaller than Set (18.1) (or Set (18.2) equivalently). The
constraint set for the output-oriented model is obtained by restricting ϕðrkÞ = ϕðkÞ
and setting θi = θf = ϕg = 1, which is also smaller than Set (18.2). The slacks-
based model is thus more general than the division parameter model.
4. System parameter model
The most restrictive case is to require all input parameters to be the same, and
ignoring all other parameters:
ðk Þ
θi = θ, 8i, k
ðk Þ ð18:5Þ
ðk Þ
θf = φg = φðrkÞ = 1, 8f , g, r, k
In this case we have the constraint set corresponding to the input-oriented system
parameter model. This set is a special case of both Sets (18.3) and (18.4), it is thus
smaller than the intersection of the two sets. The output-oriented case is ϕðrkÞ = ϕ and
θi = θf = ϕg = 1, and the corresponding constraint set is also in the intersection
of Sets (18.3) and (18.4) of the output-oriented case.
5. Independent model
ðk Þ
If we allow θf to be variables, and require them to be the same, as those defined
ðk Þ
for θi in the division parameter model, that is,
468 18 Epilogue
ðk Þ ðk Þ
θi = θf = θð k Þ , 8i, f , k
ð18:6Þ
φðrkÞ = φðgkÞ = 1, 8r, g, k
then we have a constraint set the same as that of the input-oriented independent
model, in which the efficiency of every division is measured independently. The
feasible region of this model, defined by Set (18.6), is larger than that of the division
ðk Þ
parameter model, defined by Set (18.4), because θf now lies in a range, rather than
being fixed at 1. The constraint set for the output-oriented model can be defined
similarly.
6. Directional distance function model
The conventional distance function model, with the feasible region defined in Set
(18.1), searches for a benchmark point from the origin moving along the direction
(X0, Y0, Z0) at the rate specified by the parameters associated with the factors. The
directional distance function model searches for a benchmark point from the position
of the DMU being evaluated along a pre-specified direction (d, d, d, d) at the rate
defined by the parameters associated with the direction. Its constraint set is:
n
ðk Þ ðk Þ ðk Þ ðk Þ ðkÞ
λj X ij ≤ X i0 - ηi di , 8i, k
j=1
n p p
ðk Þ ða,k Þ ða,k Þ ðk Þ ðk Þ
λj Z fj ≤ Z f 0 - ηf d f , 8f , k
j=1 a=1 a=1
n
ðk Þ
p
ðk,bÞ
p
ðk,bÞ ðk Þ ð18:7Þ
λj Z gj ≥ Z g0 þ ηðgkÞ d g , 8g, k
j=1 b=1 b=1
n
λj Y rj ≥ Y r0 þ ηðrkÞ d r , 8r, k
j=1
ηi , ηf , ηðgkÞ , ηðrkÞ , λj ≥ 0, 8i, f , g, r, j, k
If we specify
ðk Þ ðk Þ
d i = X i0 , 8i, k
p
ðk Þ ða,k Þ
df = Zf 0 , 8f , k
a=1
ðk Þ ðk Þ ð18:8Þ
d g = d r = 0, 8g, r, k
ðk Þ ðk Þ
ηi = 1 - θ i , 8i, k
ðk Þ ðk Þ
ηf = 1 - θ f , 8f , k
18.2 Which Model to Use 469
Fig. 18.1 Generality of

some representative models DDF
SBM
IND
DIV SYS FAC
then we have a constraint set the same as that of the input-oriented slacks-based
ðk Þ
model, which is Set (18.2), with ϕg and ϕðrkÞ set to 1. Similarly, if we specify
ðk Þ ðk Þ ðk Þ ðk,bÞ ðk Þ ðk Þ ðk Þ
d i = df = 0, dg = pb = 1 Z g0 , dr = Y r0 , ηðgkÞ = ϕg - 1, and ηðrkÞ = ϕðrkÞ - 1,
then we have the constraint set the same as that of the output-oriented slacks-based
model. This implies that the directional distance function model is more general than
the slacks-based model, and Set (18.2) is contained in Set (18.8).
Figure 18.1 shows the inclusion relationship of the constraint sets defined above,
where DDF, SBM, FAC, IND, DIV, and SYS denote a directional distance function
model, slacks-based measure model, factor parameter model, independent model,
division parameter model, and system parameter model, respectively. Larger con-
straint sets yield smaller objective values in the minimization case, and larger
objective values in the maximization case. However, since the objective functions
of these models are not the same, their objective values are not comparable.
Figure 18.1 thus only shows the generality of different models.
18.2 Which Model to Use
While the whole-unit model gives a general picture of the performance of a DMU, all
studies indicate that a network analysis is necessary in order to obtain reliable results
whenever the data is available. The question is which of the network DEA models
should be used.
Basically, the answer depends on the assumptions and purposes of the study.
Network DEA models can be roughly classified as independent, ratio-form, distance
function, and slacks-based. The feasible regions corresponding to these models have
inclusion relationships, as discussed in the preceding section, in that the directional
distance function model is the most general, and the system parameter model is the
most restrictive. If these models have the same objective function, then the most
general one has the lowest efficiency measure and the most restrictive one has the
470 18 Epilogue
highest when the objective is minimization. Normally higher efficiency measures are
more favorable to the DMUs being evaluated, while they are less favorable to the
decision maker, due to its low discriminating power in judging whether one DMU is
better than another. Nevertheless, unless two models have the same objective
function, their objective values are actually not comparable.
Of these classes of models, the independent model is a very straightforward one,
which measures the efficiency of each division in the network separately. A DMU is
considered efficient only if all its component divisions are efficient. The advantage
of this class of model is that the efficiency of every division is correctly measured,
where an efficiency score is considered correct if it is the largest one. Its disadvan-
tage is that it fails to see the wood for the trees. Although it shows the efficiency of
every division, the overall performance of the system is not clear. One way to
overcome this weakness is to specify a set of weights to calculate a weighted average
of the division efficiencies to represent the overall efficiency of the system. How-
ever, this efficiency aggregation method has a problem, which will be discussed in
the next section. Another disadvantage is that each division of the system must be
enumerated for all DMUs in measuring efficiencies, which is time consuming and
troublesome.
The ratio-form and distance function models are the dual of each other, and are
essentially the same. However, they provide different information in addition to the
efficiency of the system. For the ratio-form model the efficiencies of all divisions are
also obtainable. Moreover, there exist relationships between the system and division
efficiencies for structured systems, i.e., systems with the structures discussed in
Chaps. 10, 11, 12, 13, 14, and 15. The relationship helps identify the bottleneck
divisions, whose efficiencies have larger effects than other divisions on that of the
system. The same effort devoted to the operations of these divisions brings the
system efficiency to a higher level than would be seen if it was devoted to other
divisions. From this standpoint the ratio-form model is better than the independent
model. However, since alternative solutions may exist in the ratio-form model, the
division efficiencies calculated from the solution obtained in measuring the effi-
ciency of different DMUs may be inconsistent for comparison purposes. It turns out
that the model may need to be repeated for all divisions, in the worst case, in order to
ensure that the calculated division efficiencies are comparable. The extra information
obtained by the distance function model is a set of target values for all factors for
inefficient DMUs to achieve to become efficient. When the ratio-form and distance
function models are used together, both the division efficiencies and targets can be
obtained.
The major disadvantage of the ratio-form model (and the distance function model
as its dual) is that the calculated efficiency scores do not reflect the proper order of
the performance of weakly efficient DMUs and those using a weakly efficient DMU
as the target to measure efficiency. A weakly efficient DMU has an efficiency score
very close to 1 (in the form of 1-cε, where c is a constant and ε is a small
non-Archimedean number). However, if the difference between the observed and
target values for all factors are considered equally important, then the efficiency
score of the weakly efficient DMUs may drop to a level lower than that of some
18.3 Road Ahead 471
inefficient DMUs. This class of model should thus be used when rankings are not the
major concern of the study.
The slacks-based model has the advantages of both the ratio-form and distance
function models. It is able to measure the efficiency of the divisions, and to show the
targets for making improvements in one model. Moreover, it provides proper
efficiency measures for weakly efficient DMUs. However, this model is not without
weaknesses. It is noted in empirical studies that the efficiency score of a unit can be
unexpectedly low, due to a large target value in one output, which raises doubts as to
whether the efficiency score reflects the true performance of the unit.
In sum, since each model has advantages and disadvantages, it is advised to use
more than one model in real world applications when there are no specific assump-
tions or purposes for guidance. The results from different models provide more
information to draw more confident conclusions regarding the performance of
the DMUs.
18.3 Road Ahead
The preceding section points out the advantages and disadvantages of different
efficiency measurement models. Since a model that does not have weaknesses is
preferable, it is thus necessary to find a more appropriate way to calculate the
efficiency for weakly efficient DMUs for the ratio-form model (and its dual
model), and a more appropriate way to calculate the efficiency for DMUs with one
extremely inefficient factor for the SBM model. This involves introducing new
ideas, which is not an easy task.
The different structures of network systems that have been identified in this book
include basic series, general series, parallel, and hierarchical. The purpose of iden-
tifying a structure type is to make modelling easier, and to explore the properties
possessed by the system and division efficiencies. For example, once the numbers of
inputs, outputs, and DMUs are specified for the conventional whole-unit system, a
model for measuring the efficiency of a set of DMUs can then be formulated, and a
computer program be developed. When the data of all DMUs has been obtained,
then their efficiencies can be calculated with one stroke of a key on a computer
keyboard. However, the types of structure that have been identified are not exhaus-
tively inclusive, in that not every network structure can be classified into one of
them. If more structure types are explored, and all possible structures are included,
then measuring efficiency will be easier, and the efficiency scores will be more
informative.
One purpose of developing network DEA models is to find more appropriate
ways to measure the efficiency of DMUs. Another purpose, which is more impor-
tant, is to explore the relationship between the system and division efficiencies.
There are at least three aspects that require further investigation, data, returns to
scale, and time. As revealed by its name, the basis of data envelopment analysis is
data. Different types of data have been studied for the whole-unit system, for
472 18 Epilogue
example, negative, ordinal, qualitative, interval, probabilistic, and fuzzy. How to

apply the methods that have been used to measure the efficiency of the whole-unit
system to network systems needs further study, and the issue can be discussed at two
levels. At the lower level one question is whether the properties that hold for the
conventional real-valued data still hold for the special types of data. For example, a
property of the general series system when the data is real-valued is that the system
efficiency lies between the product and weighted average of the division efficiencies.
Whether this property holds or not when the data is of other types is not known. At
the higher level one question is how efficiency will be affected when the data is less
precise, and how the system efficiency will be affected by the division efficiencies
when the network structure becomes more complicated. The former case indicates
how important it is to collect the data with a greater precision.
Returns to scale is another topic that is worth investigating in this context. The
divisions in a network may have different returns to scale, and the system as a whole
unit has a specific one, such as increasing, constant, and decreasing, depending on
production types and assumptions. How the aggregation of the production technol-
ogies of individual divisions leads to that of the system for different network
structures requires theoretical justification and empirical verification, thus presenting
directions for future research.
Another issue that has not been addressed in network DEA is time. If time is
considered as a factor in measuring efficiency, then divisions that operate faster have
higher efficiencies, and those operate slower have lower efficiencies. For a network
system with some divisions operating fast and some slow, the effects of these on the
performance of the system is dependent on the structure of the system. Many
network models use the aggregation of the division efficiencies to represent the
system efficiency, which may not be appropriate. Consider a simple assembly
system of three divisions, where two operate in parallel to produce intermediate
products for the third to assemble to become the final product. If one of the two
parallel divisions operate relatively fast and the other relatively slow to result in a
high and a low efficiency, then the third division must wait for the slow division to
finish the work to be able to start its operations. The operation of the whole system is
thus relatively slow, due to the slow division. That is, the impact of the slow division
dominates the performance of the system. When using the aggregation of division
efficiencies to represent the system efficiency in this case, the efficiency of the fast
division will dilute that of the slow division to result in a moderate efficiency of the
system, which is obviously inappropriate. This time interdependence relationship
between divisions must be properly considered to find the correct relationship
between the system and division efficiencies.
Finally, real world problems are complicated. They have been either simplified or
separated into smaller ones to suit existing models. The existing models thus need to
be incorporated with other methods to be able to solve the problems in their complete
form. This is a challenge to the development of network DEA models.
Index
A Cost efficiency, 358, 461

Absolute-directional method, 140 Cross efficiency, 38, 40
Additive model, 66–69, 82, 85, 275, 285, 308,
414, 415
Allocative efficiency, 8, 9, 11, 62, 245 D
Assurance region, 35–37 Directional distance function, 43, 59–62,
83–85, 111, 129, 163, 176, 178, 180,
276, 277, 310, 343, 368, 403–406, 414,
B 468–469
Bargain game, 230 Disposability, 46, 91, 98, 99, 105, 106, 110,
BCC model, 26–33, 37, 38, 40, 66, 67, 89, 91, 127, 131, 238, 257, 270, 404
92, 95, 98, 100, 101, 104–107, 109, 110, Distance function, 2, 12, 43–62, 65, 69, 70, 82,
116, 118, 120, 121, 140, 143, 144, 222, 91, 158, 163–167, 172, 176, 184, 190,
245, 285 219, 231–242, 246, 251, 270–277, 286,
Benchmark, 6, 8, 10–12, 24, 25, 29, 33, 43, 291, 305–310, 321, 335, 338–343, 358,
48–53, 55–59, 61, 62, 65–73, 75, 77, 363, 370, 386, 391, 396, 400–406, 413,
82–84, 116, 117, 120, 169, 465, 468 469–471
Bounded adjusted measure (BAM), 85 Dynamic model, 396, 401, 408, 413
C E
Carryover, 396, 397, 401, 403, 405, 410–415, Efficiency aggregation, 14, 223, 231–236, 244,
460 245, 254, 255, 261, 263–264, 285,
Categorical-data, 145 296–298, 300–301, 312, 313, 315, 316,
CCR model, 20–26, 28, 33, 37, 70, 71, 73, 97, 319, 320, 328, 333, 336, 344–346, 384,
98, 103, 121, 122, 183, 187, 188, 222, 399, 457, 461, 470
329, 347, 357 Efficiency decomposition, 14, 172, 223–232,
Common weight, 37, 38, 357 234, 237, 241, 244, 255, 260,
Cone ratio, 36 264–270, 272, 292–297, 301–306,
Congestion, 89, 90, 104–112 313, 316, 318, 319, 328, 333, 336, 345,
Connected model, 190–199, 206–210, 212, 215 346, 350, 367, 377, 424, 432, 435, 453,
Cooperative game, 263, 461 457, 460
Cooperative model, 199–202, 213–215, 423, Envelopment form, 22, 24, 31, 32, 48, 50, 54,
428, 443, 449 56, 66, 72, 120, 186, 188–189, 191–195,
https://doi.org/10.1007/978-3-031-27593-7
474 Index
197–201, 203, 206, 207, 211, 213, 270, M

286, 330, 334, 457 Malmquist productivity index (MPI), 163–176,
Envelopment model, 192, 199, 211, 238, 244, 178, 179, 285, 286
246, 276, 331, 334 Measure of inefficiency proportion (MIP), 68,
Extension principle, 155, 157 72, 74
External evaluation, 447–454, 460 Membership function, 153–155, 158
Multi-component system, 327, 328, 331, 334,
347, 352
F Multi-function system, 325, 331, 332, 347, 352
Feedback, 251–257, 384, 419, 432, 450, 457, 461 Multiplicative model, 90–98, 100, 111
Free disposal hull (FDH), 62, 90, 98–104, 111, Multiplier form, 21, 24, 31, 32, 67, 185,
159, 415, 466 187–188, 193–194, 196–197, 200–201,
Fuzzy data, 153–158, 215, 245, 321, 336, 357 203, 207, 210, 213, 223, 241, 272, 281
Fuzzy number, 137, 145, 153–155, 357 Multiplier model, 246
Multi-stage system, 15, 291–321, 331, 396
G
Game-theoretic approach, 230–231 N
Gauge function, 62 Negative data, 138–141, 158
Global Malmquist productivity index, 173–176, Non-Archimedean number, 20, 34, 48, 49, 54,
180, 231 55, 69, 72, 92, 99, 104, 106, 186, 224,
233, 234, 237, 257, 271, 306, 330, 334,
339, 342, 400, 401, 420, 425, 431, 432,
H 470
Hierarchical structure, 15, 325, 326, 349, 359, Non-cooperative game, 286
461 Non-discretionary factor, 117, 118, 120–122,
Hyperbolic model, 128–130, 413 132
Nonparametric approach, 2, 6, 11, 146
Nonzero sum game, 230
I
Imprecise data, 85, 137, 153, 158, 357
Independent model, 187–190, 192, 193, 195, O
203–208, 215, 219–223, 225, 226, 230, Ordinal data, 141–143, 158
231, 234, 245, 261, 265, 298, 320, 365, Output model, 12, 23–26, 30–33, 35, 38, 61, 65,
468–470 66, 72–74, 77, 80, 83–85, 99, 100, 105,
Input model, 20–23, 25, 27–34, 50, 61, 65, 118–120, 122, 128, 143, 187, 195, 226,
69–72, 77, 79, 83, 99, 100, 105, 116–119, 234, 239, 272, 402, 457
121, 122, 148, 151, 155, 185–188, Output possibility set, 45, 46, 54, 58
195–197, 203, 226, 232, 239, 340, 455
Input possibility set, 44, 45, 47–49, 54
Internal evaluation, 447, 448, 454–460 P
Intertemporal effect, 395, 415 Parallel structure, 14, 325, 326, 355, 369, 372,
Interval data, 137, 151–153, 157, 158, 285 377, 380, 388, 389, 435, 441
Isocost, 8, 10 Parametric approach, 2, 6, 9, 11, 146
Isoquant, 2, 7–9, 23, 44, 45, 48, 70, 97, 101, Pareto efficiency, 20, 34, 54
109, 110, 116, 117, 123, 260 Pareto optimality, 20
Isorevenue, 10 Price efficiency, 8
Isotonic, 158 Product transformation curve, 10, 11, 25, 45,
54, 73, 119, 127, 259, 260
Production delay, 395, 401–402
L Production frontier, 5–12, 21, 22, 24–26, 28,
Linkage efficiency, 15, 420, 422, 423, 429, 432, 30, 32, 33, 44, 49, 59, 61, 65, 66, 82, 83,
433, 435, 437, 441–443, 445, 454 89, 90, 94–97, 100–104, 110, 124, 125,
Luenberger productivity index (LPI), 163, 138, 144, 147, 164, 167, 168, 220, 228,
176–179 229, 259, 260, 327, 329, 351
Index 475
Production function, 2–7, 9, 43, 44, 46, 91, 97, Series structure, 14, 268, 291–298, 300–302,
120, 164, 167 304, 314, 325, 382, 386, 388, 391
Production possibility set, 43–48, 54, 55, 59, Shared input, 251, 281–285, 287, 325,
62, 66, 70, 75, 89–91, 94, 95, 98–102, 334–346, 357–359, 405
112, 124, 126, 131, 143, 190, 191, 208, Shortage function, 62, 176
212, 236, 237, 270, 278, 415, 465, 466 Slacks-based measure (SBM), 12, 65, 72, 74,
Projection point, 80, 81, 85, 93, 103, 107, 108, 79–81, 84, 85, 131, 158, 189, 194, 199,
116, 128–130, 138, 169, 192, 195, 198, 201, 212, 215, 246, 321, 410, 414, 415,
200, 207, 211, 222, 238 420, 421, 423, 425, 431, 432, 435, 444,
Pure technical efficiency, 104, 106–109, 111, 112 448, 453–455, 457–460, 469, 471
Static model, 395–397, 409
Stochastic data, 85, 146–151, 215
Q Stochastic frontier approach, 146
Qualitative data, 142–145, 158 Strong efficient, 21, 116, 118, 119, 145, 275,
Quasi-fixed input, 413 447
Super efficiency, 38, 39, 85, 426
Supply chain, 112, 246, 261, 285, 314, 321,
R 445, 461
Radial measures, 8–10, 43, 65, 66, 69, 82, 83, Supporting hyperplane, 30, 69, 82, 93
103, 132, 158, 280, 419, 420, 444, 448
Range adjusted measure (RAM), 68
Range-directional method, 140 T
Rankings, 36–40, 65, 83, 84, 140, 146, 246, Technical efficiency, 30, 33, 50, 74, 97, 111,
260, 287, 471 172, 226–229, 392, 413
Ratio model, 48, 52, 54, 56, 78–82, 121, 238, Törnqvist index, 165
239, 276, 341 Translation invariant, 68, 139
Ray unboundedness, 47, 95, 102 Two-level mathematical programming, 152,
Reference point, 50, 52, 53, 56, 58, 75 156
Relational model, 195–201, 210–215, 224, 226, Two-person game, 261
266, 274, 279, 287, 326, 348, 357, 363, Two-stage system, 14, 15, 219–246, 251,
365, 368, 371, 373–375, 379, 381, 390, 255, 256, 258, 266, 270, 272, 275,
420, 424, 430, 443, 444 281, 283, 285, 287, 291–293, 365,
Returns to scale, 2, 21, 25–33, 37–40, 47–59, 369, 445, 461
66, 68, 70, 72, 75–77, 79–82, 89–98,
101–104, 111, 116, 118, 121–122, 127,
139, 163, 167, 172–174, 178, 179, 190, U
191, 193, 194, 196, 208, 220, 222–230, Undesirable factors, 115, 123–132, 244, 277,
232, 234–236, 240, 243, 252, 258, 270, 279, 372, 378, 379, 386, 387, 403, 404,
271, 278, 281, 293, 299, 300, 305–307, 407, 411
313, 321, 326, 331, 335, 347, 364–366, Unit invariant, 66, 68, 139
368, 370, 374, 375, 379, 381, 384, 386,
392, 397, 400, 402, 404, 407, 415, 420,
423, 425, 431, 432, 443, 449, 453, V
455–457, 461, 466, 471, 472 Value-based model, 321, 359, 412
Revenue efficiency, 415
Reversal link, 291, 314–321, 346
Russell measure, 65, 69–78, 85 W
Weak disposability, 90, 105–110, 112, 123,
126–131, 133, 404, 414
S Weakly efficient, 21, 23, 48, 53, 65, 69, 71, 83,
Scale efficiency, 28, 30, 33, 50, 52, 56, 97, 98, 84, 92, 99, 101, 116–119, 145, 246, 275,
103, 106, 107, 109, 122, 172, 226–229, 277, 420, 471
245 Window analysis, 33

Network Data Envelopment Analysis

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Network Data Envelopment Analysis

Uploaded by

Copyright:

Available Formats

International Series in

Operations Research & Management Science

Editorial Board Members

Network Data Envelopment

ISSN 0884-8289 ISSN 2214-7934 (electronic)

and Extension Education), Shih-Pin Chen of National Chung Cheng University

Tainan, Taiwan Chiang Kao

efﬁciency of a network system is whether the intermediate products should be

Tainan, Taiwan Chiang Kao

4 Slacks-Based Efﬁciency Measures . . . . . . . . . . . . . . . . . . . . . . . . . . 65

7 Special Types of Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 137

9.7 Supplementary Literature . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 215

12.6 Reversal Links . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 314

15.2.3 Directional Distance Function: Whole Unit . . . . . . . . . 403

Chiang Kao is a University Chair Professor of Industrial and Information Manage-

Performance evaluation is an important task in management, to better understand the

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2023 1

1.1 History of Network DEA

In physical sciences theoretical standards showing the relationship between several

1.2 Basic Ideas of Efﬁciency Measurement

In economic theory a production function is a mathematical equation showing the

Fig. 1.1 Production Y

yields a = 2.5093 and b = 5.9807, with the regression curve shown by OP in

1.3 Multi-Input Case

Fig. 1.2 A production Y

Fig. 1.3 Isoquant and input X2

1.4 Multi-Output Case

In contrast to using multiple inputs to produce single output, another situation is

Fig. 1.4 A production O Y2

Fig. 1.5 Product Y2

1.5 Whole-Unit Analysis

Fig. 1.6 Whole-unit

1.6 Network Analysis

division k. The same intermediate product f produced by different divisions for

Fig. 1.7 Network

Fig. 1.8 The operation of X i( k ) , i ∈ I ( k )

Z (f1,k ) Z g(k ,1)

1.7 Supplementary Literature

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2023 19

2.1 CCR Model

Charnes et al. (1978) proposed a fractional programming model, commonly referred

2.1.1 Input Model

Table 2.1 CCR efﬁciency of DMU Input X Output Y Efﬁciency

Fig. 2.1 CCR efﬁciency Y

Table 2.2 CCR efﬁciency of DMU Input X1 Input X2 Output Y Efﬁciency

Fig. 2.2 CCR efﬁciency X2

Consider another case, where ﬁve DMUs, A, B, C, D, and E, applying different

2.1.2 Output Model

Table 2.3 CCR efﬁciency of DMU Input X Output Y1 Output Y2 Efﬁciency

Fig. 2.3 CCR efﬁciency Y2

2.2 BCC Model

2.2.1 Input Model

Model (2.9) has a dual, which can be formulated as:

Fig. 2.4 Efﬁciency measurement under variable returns to scale

E 0 = max: ðu1 Y 0 - u0 Þ=X 0

2.2.2 Output Model

Table 2.5 Summary of the BCC models in different forms

(2.7), is the convexity constraint of nj = 1 λj = 1. Table 2.5 summarizes the BCC

2.3 Restrictions on Multipliers

Fig. 2.5 Frontiers adjusted X2

In this form the importance of each input/output factor is expressed in relation to

Fig. 2.6 Frontiers adjusted

An issue closely related to restrictions on multipliers is ranking. The DEA technique

At optimality, the efﬁciency of DMU j is calculated as